From e3d68ccc3714175a12854428f34adbb69f09474f Mon Sep 17 00:00:00 2001
From: cai <cai@nbcai.cc>
Date: Sat, 08 Aug 2026 18:42:58 +0800
Subject: [PATCH] chore: bind dynamic metadata test candidate

---
 src/main.rs | 1210 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 files changed, 1,178 insertions(+), 32 deletions(-)

diff --git a/src/main.rs b/src/main.rs
index 05ba69b..d357e4a 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -1,7 +1,10 @@
+mod asr_realtime;
 mod audio;
 mod service;
 
 use std::{
+    borrow::Cow,
+    collections::HashSet,
     env, fs,
     path::{Path, PathBuf},
     sync::{
@@ -12,6 +15,9 @@
 };
 
 use anyhow::{Context, Result, anyhow};
+use asr_realtime::{
+    AudioIngressMetadata, RealtimeAsrConfig, RealtimeAsrOutcome, RealtimeAsrUpload,
+};
 use audio::{AudioDiagnostics, load_pre_recorded_frames};
 use base64::{Engine as _, engine::general_purpose};
 use futures_util::StreamExt;
@@ -24,23 +30,35 @@
     options::TrackPublishOptions,
     prelude::{
         DataPacket, LocalAudioTrack, LocalTrack, ParticipantIdentity, RemoteAudioTrack,
-        RemoteTrack, Room, RoomEvent, RoomOptions,
+        RemoteParticipant, RemoteTrack, Room, RoomEvent, RoomOptions,
     },
 };
 use reqwest::Client;
 use serde::{Deserialize, Serialize};
 use serde_json::json;
-use tokio::time::{sleep, sleep_until};
-use tokio::{sync::mpsc::UnboundedReceiver, task::JoinHandle};
+use sha2::{Digest, Sha256};
+use tokio::time::{sleep, sleep_until, timeout};
+use tokio::{
+    sync::{mpsc, mpsc::UnboundedReceiver, watch},
+    task::JoinHandle,
+};
 use tracing::{info, warn};
 
 const USER_AUDIO_SAMPLE_RATE_HZ: u32 = 48_000;
 const USER_AUDIO_NUM_CHANNELS: u16 = 1;
+const INBOUND_AUDIO_QUEUE_CAPACITY: usize = 100;
+const INBOUND_AUDIO_DROP_LOG_INTERVAL: u64 = 100;
+const AUDIO_DRAIN_STOP_GRACE: Duration = Duration::from_secs(2);
 const DEFAULT_BOT_AUDIO_PROFILE: &str = "pcm-16k";
 const LIVEKIT_48K_SAMPLE_RATE_HZ: u32 = 48_000;
 const PCM_16K_SAMPLE_RATE_HZ: u32 = 16_000;
 const BOT_NUM_CHANNELS: u16 = 1;
 const TRACK_NAME: &str = "bot-main-audio";
+const STREAM_TIMING_VERSION: u32 = 1;
+const STREAM_TIMING_FIRST_SEGMENT: u64 = 1;
+const STREAM_TIMING_FIRST_CHUNK: u64 = 1;
+const STREAM_TIMING_MAX_ELAPSED_MS: u64 = 5_000;
+const STREAM_TIMING_SOURCE: &str = "stream_anchor_monotonic";
 
 #[tokio::main(flavor = "multi_thread")]
 async fn main() -> Result<()> {
@@ -270,6 +288,11 @@
     runtime_turn_bridge_url: Option<String>,
     runtime_turn_bridge_token: Option<String>,
     runtime_turn_bridge_mode: String,
+    runtime_asr_stream_enabled: bool,
+    runtime_asr_stream_url: Option<String>,
+    runtime_asr_realtime_enabled: bool,
+    runtime_asr_realtime_url: Option<String>,
+    runtime_asr_realtime_chunk_duration_ms: u64,
     runtime_turn_artifact_dir: Option<String>,
     runtime_session_nonce: Option<String>,
     user_audio_observer_enabled: bool,
@@ -332,6 +355,11 @@
     bridge_url: Option<String>,
     bridge_token: Option<String>,
     bridge_mode: String,
+    asr_stream_enabled: bool,
+    asr_stream_url: Option<String>,
+    asr_realtime_enabled: bool,
+    asr_realtime_url: Option<String>,
+    asr_realtime_chunk_duration_ms: u64,
     artifact_dir: Option<String>,
     runtime_session_nonce: Option<String>,
     audio_debug_dump_dir: Option<String>,
@@ -343,6 +371,11 @@
             bridge_url: config.runtime_turn_bridge_url.clone(),
             bridge_token: config.runtime_turn_bridge_token.clone(),
             bridge_mode: config.runtime_turn_bridge_mode.clone(),
+            asr_stream_enabled: config.runtime_asr_stream_enabled,
+            asr_stream_url: config.runtime_asr_stream_url.clone(),
+            asr_realtime_enabled: config.runtime_asr_realtime_enabled,
+            asr_realtime_url: config.runtime_asr_realtime_url.clone(),
+            asr_realtime_chunk_duration_ms: config.runtime_asr_realtime_chunk_duration_ms,
             artifact_dir: config.runtime_turn_artifact_dir.clone(),
             runtime_session_nonce: config.runtime_session_nonce.clone(),
             audio_debug_dump_dir: config.audio_debug_dump_dir.clone(),
@@ -374,6 +407,32 @@
                 .as_deref()
                 .is_some_and(|value| value.trim_end_matches('/').ends_with("/stream"))
     }
+
+    fn is_asr_stream_ready(&self) -> bool {
+        self.asr_stream_enabled
+            && self
+                .asr_stream_url
+                .as_ref()
+                .is_some_and(|value| !value.is_empty())
+            && self
+                .bridge_token
+                .as_ref()
+                .is_some_and(|value| !value.is_empty())
+            && self
+                .runtime_session_nonce
+                .as_ref()
+                .is_some_and(|value| !value.is_empty())
+    }
+
+    fn realtime_asr_config(&self) -> RealtimeAsrConfig {
+        RealtimeAsrConfig {
+            enabled: self.asr_realtime_enabled,
+            url: self.asr_realtime_url.clone(),
+            runtime_token: self.bridge_token.clone(),
+            runtime_session_nonce: self.runtime_session_nonce.clone(),
+            chunk_duration_ms: self.asr_realtime_chunk_duration_ms,
+        }
+    }
 }
 
 impl Config {
@@ -395,6 +454,14 @@
             runtime_turn_bridge_token: optional_env("CV_RUNTIME_TURN_BRIDGE_TOKEN"),
             runtime_turn_bridge_mode: env::var("CV_RUNTIME_TURN_BRIDGE_MODE")
                 .unwrap_or_else(|_| "json".to_string()),
+            runtime_asr_stream_enabled: bool_env("CV_RUNTIME_ASR_STREAM_ENABLED", false),
+            runtime_asr_stream_url: optional_env("CV_RUNTIME_ASR_STREAM_URL"),
+            runtime_asr_realtime_enabled: bool_env("CV_RUNTIME_ASR_REALTIME_ENABLED", false),
+            runtime_asr_realtime_url: optional_env("CV_RUNTIME_ASR_REALTIME_URL"),
+            runtime_asr_realtime_chunk_duration_ms: u64_env(
+                "CV_RUNTIME_ASR_REALTIME_CHUNK_DURATION_MS",
+                200,
+            ),
             runtime_turn_artifact_dir: optional_env("CV_RUNTIME_TURN_ARTIFACT_DIR"),
             runtime_session_nonce: optional_env("CV_RUNTIME_SESSION_NONCE"),
             user_audio_observer_enabled: bool_env("CV_ENABLE_USER_AUDIO_OBSERVER", true),
@@ -605,6 +672,31 @@
     println!("{payload}");
 }
 
+fn emit_anchored_activity(
+    call_id: &str,
+    trace_id: &str,
+    turn_id: &str,
+    event_name: &str,
+    marker: &RuntimeTurnStreamTimingMarker,
+    extension: serde_json::Value,
+) {
+    let payload = json!({
+        "type": "cv_activity",
+        "callId": call_id,
+        "traceId": trace_id,
+        "turnId": turn_id,
+        "eventName": event_name,
+        "eventWallTimeMs": current_time_millis(),
+        "serverDeltaMs": marker.server_delta_ms,
+        "serverDeltaSource": STREAM_TIMING_SOURCE,
+        "result": "ok",
+        "reasonCode": null,
+        "retryable": null,
+        "extension": marker.extension_with(extension),
+    });
+    println!("{payload}");
+}
+
 fn spawn_user_audio_observer(
     events: UnboundedReceiver<RoomEvent>,
     config: &Config,
@@ -617,6 +709,7 @@
     let enabled = config.user_audio_observer_enabled;
     let simple_vad_enabled = config.simple_vad_enabled;
     let simple_vad_config = config.simple_vad_config.clone();
+    let user_participant_identity = config.user_participant_identity.clone();
     let turn_bridge_config = TurnBridgeConfig::from_config(config);
 
     tokio::spawn(async move {
@@ -638,6 +731,7 @@
             turn_bridge_config,
             http,
             sink,
+            user_participant_identity,
         )
         .await;
     })
@@ -653,6 +747,7 @@
     turn_bridge_config: TurnBridgeConfig,
     http: Client,
     sink: Arc<BotAudioOutputSink>,
+    user_participant_identity: Option<String>,
 ) {
     info!(
         call_id = %call_id,
@@ -676,6 +771,15 @@
                 publication: _,
                 participant,
             } => {
+                if user_participant_identity
+                    .as_deref()
+                    .is_some_and(|expected| participant.identity().to_string() != expected)
+                {
+                    warn!(call_id = %call_id, trace_id = %trace_id,
+                        metadata_status = "wrong_participant",
+                        "runtime helper ignored non-user audio participant");
+                    continue;
+                }
                 let participant_alias = redact(&participant.identity().to_string());
                 let track_sid_alias = redact(&track.sid().to_string());
                 let track_name = track.name();
@@ -701,6 +805,7 @@
                     turn_bridge_config.clone(),
                     http.clone(),
                     sink.clone(),
+                    participant,
                 );
             }
             RoomEvent::TrackSubscribed {
@@ -751,6 +856,7 @@
     call_id: &str,
     trace_id: &str,
     turn: FinishedSpeechTurn,
+    realtime_asr_result_ref: Option<String>,
 ) {
     let turn_pipeline_started_at = Instant::now();
     if !bridge_config.is_ready() {
@@ -807,9 +913,13 @@
                     "turnArtifactBytes": byte_size,
                     "frameCount": turn.frame_count,
                     "sampleCount": turn.sample_count,
-                    "endReason": turn.end_reason,
+                    "endReason": turn.end_reason.as_str(),
                 }),
             );
+            let asr_result_ref = match realtime_asr_result_ref {
+                Some(value) => Some(value),
+                None => request_asr_result_ref(http, bridge_config, call_id, trace_id, &turn).await,
+            };
             if bridge_config.is_stream_mode() {
                 match request_turn_bridge_stream(
                     http,
@@ -820,6 +930,7 @@
                     &turn,
                     &path_ref,
                     byte_size,
+                    asr_result_ref.as_deref(),
                     turn_pipeline_started_at,
                 )
                 .await
@@ -883,12 +994,17 @@
                 &turn,
                 &path_ref,
                 byte_size,
+                asr_result_ref.as_deref(),
                 turn_pipeline_started_at,
             )
             .await
             {
                 Ok(outcome) => {
+                    let mut published_device_outputs = HashSet::new();
                     for output in &outcome.device_outputs {
+                        if !should_publish_device_output(&mut published_device_outputs, output) {
+                            continue;
+                        }
                         if let Err(error) = sink
                             .publish_device_output(call_id, trace_id, &turn.turn_id, output)
                             .await
@@ -1004,6 +1120,251 @@
     Ok((path_ref, byte_size))
 }
 
+async fn request_asr_result_ref(
+    http: &Client,
+    bridge_config: &TurnBridgeConfig,
+    call_id: &str,
+    trace_id: &str,
+    turn: &FinishedSpeechTurn,
+) -> Option<String> {
+    if !bridge_config.is_asr_stream_ready() {
+        return None;
+    }
+    match request_asr_stream(http, bridge_config, call_id, trace_id, turn).await {
+        Ok(Some(asr_result_ref)) => {
+            emit_activity(
+                call_id,
+                trace_id,
+                Some(&turn.turn_id),
+                "asr_stream_ref_ready",
+                "ok",
+                None,
+                None,
+                json!({
+                    "asrResultRefPresent": true,
+                    "format": "pcm_s16le",
+                    "sampleRate": 16000,
+                    "channels": 1,
+                }),
+            );
+            Some(asr_result_ref)
+        }
+        Ok(None) => None,
+        Err(error) => {
+            warn!(
+                call_id = %call_id,
+                trace_id = %trace_id,
+                turn_id = %turn.turn_id,
+                error = %safe_error(&error.to_string()),
+                "runtime helper asr_stream_failed_fallback"
+            );
+            emit_activity(
+                call_id,
+                trace_id,
+                Some(&turn.turn_id),
+                "asr_stream_fallback",
+                "skipped",
+                Some("ASR_STREAM_INTERRUPTED"),
+                Some(true),
+                json!({
+                    "fallbackReason": "asr_stream_request_failed",
+                    "fallbackStage": "asr_stream",
+                }),
+            );
+            None
+        }
+    }
+}
+
+async fn request_asr_stream(
+    http: &Client,
+    bridge_config: &TurnBridgeConfig,
+    call_id: &str,
+    trace_id: &str,
+    turn: &FinishedSpeechTurn,
+) -> Result<Option<String>> {
+    let started_at = Instant::now();
+    let ndjson = build_asr_stream_ndjson(call_id, trace_id, turn, bridge_config)?;
+    let response = http
+        .post(bridge_config.asr_stream_url.as_deref().unwrap_or_default())
+        .header("Content-Type", "application/x-ndjson")
+        .header(
+            "X-CV-Runtime-Token",
+            bridge_config.bridge_token.as_deref().unwrap_or_default(),
+        )
+        .header("X-CV-Call-Id", call_id)
+        .header("X-CV-Trace-Id", trace_id)
+        .header(
+            "X-CV-Runtime-Session-Nonce",
+            bridge_config
+                .runtime_session_nonce
+                .as_deref()
+                .unwrap_or_default(),
+        )
+        .body(ndjson)
+        .send()
+        .await
+        .context("failed to post asr stream")?;
+    let status = response.status();
+    if !status.is_success() {
+        let body_len = response.text().await.map(|body| body.len()).unwrap_or(0);
+        warn!(
+            call_id = %call_id,
+            trace_id = %trace_id,
+            turn_id = %turn.turn_id,
+            http_status = status.as_u16(),
+            body_len,
+            "runtime helper asr_stream_http_failed"
+        );
+        return Ok(None);
+    }
+    let body: RuntimeTurnCommonResult<RuntimeAsrStreamResp> = response
+        .json()
+        .await
+        .context("failed to decode asr stream response")?;
+    if body.code != 0 {
+        warn!(
+            call_id = %call_id,
+            trace_id = %trace_id,
+            turn_id = %turn.turn_id,
+            code = body.code,
+            msg_len = body.msg.as_deref().unwrap_or_default().len(),
+            "runtime helper asr_stream_common_result_failed"
+        );
+        return Ok(None);
+    }
+    let Some(data) = body.data else {
+        return Ok(None);
+    };
+    if data.status.as_deref() == Some("final") {
+        info!(
+            call_id = %call_id,
+            trace_id = %trace_id,
+            turn_id = %turn.turn_id,
+            chunk_count = data.chunk_count.unwrap_or_default(),
+            audio_bytes = data.audio_bytes.unwrap_or_default(),
+            asr_duration_ms = data.asr_duration_ms.unwrap_or_default(),
+            wall_ms = started_at.elapsed().as_millis() as u64,
+            provider = %data.provider_alias.as_deref().unwrap_or("unknown"),
+            text_len = data.text_len.unwrap_or_default(),
+            "runtime helper asr_stream_final"
+        );
+        return Ok(data.asr_result_ref);
+    }
+    emit_activity(
+        call_id,
+        trace_id,
+        Some(&turn.turn_id),
+        "asr_stream_fallback",
+        "skipped",
+        None,
+        Some(true),
+        json!({
+            "fallbackReason": data.fallback_reason,
+            "fallbackStage": data.fallback_stage,
+            "status": data.status,
+        }),
+    );
+    Ok(None)
+}
+
+fn build_asr_stream_ndjson(
+    call_id: &str,
+    trace_id: &str,
+    turn: &FinishedSpeechTurn,
+    bridge_config: &TurnBridgeConfig,
+) -> Result<String> {
+    let chunks = asr_pcm_16k_chunks(turn)?;
+    let mut seq = 1u64;
+    let mut lines = Vec::with_capacity(chunks.len() + 3);
+    lines.push(serde_json::to_string(&json!({
+        "event": "asr_stream_started",
+        "seq": seq,
+        "callId": call_id,
+        "traceId": trace_id,
+        "turnId": turn.turn_id.as_str(),
+        "tsMs": current_time_millis(),
+        "payload": {
+            "format": "pcm_s16le",
+            "sampleRate": 16000,
+            "channels": 1,
+            "runtimeSessionNonce": bridge_config.runtime_session_nonce.as_deref().unwrap_or_default(),
+            "providerHint": "volcengine",
+        }
+    }))?);
+    for (index, samples) in chunks.iter().enumerate() {
+        seq += 1;
+        let bytes = pcm_i16_to_le_bytes(samples);
+        let duration_ms = ((samples.len() as u64) * 1000 / 16_000).max(1);
+        lines.push(serde_json::to_string(&json!({
+            "event": "asr_audio_chunk",
+            "seq": seq,
+            "callId": call_id,
+            "traceId": trace_id,
+            "turnId": turn.turn_id.as_str(),
+            "tsMs": current_time_millis(),
+            "payload": {
+                "chunkSeq": index + 1,
+                "format": "pcm_s16le",
+                "sampleRate": 16000,
+                "channels": 1,
+                "durationMs": duration_ms,
+                "payloadBase64": general_purpose::STANDARD.encode(bytes),
+            }
+        }))?);
+    }
+    seq += 1;
+    lines.push(serde_json::to_string(&json!({
+        "event": "vad_speech_end",
+        "seq": seq,
+        "callId": call_id,
+        "traceId": trace_id,
+        "turnId": turn.turn_id.as_str(),
+        "tsMs": current_time_millis(),
+        "payload": {
+            "endReason": turn.end_reason.as_str(),
+            "speechDurationMs": turn.duration_ms,
+        }
+    }))?);
+    seq += 1;
+    lines.push(serde_json::to_string(&json!({
+        "event": "asr_stream_finish",
+        "seq": seq,
+        "callId": call_id,
+        "traceId": trace_id,
+        "turnId": turn.turn_id.as_str(),
+        "tsMs": current_time_millis(),
+        "payload": {
+            "finalChunkSeq": chunks.len(),
+            "audioDurationMs": turn.duration_ms,
+        }
+    }))?);
+    Ok(lines.join("\n") + "\n")
+}
+
+fn asr_pcm_16k_chunks(turn: &FinishedSpeechTurn) -> Result<Vec<Vec<i16>>> {
+    if turn.samples.is_empty() {
+        return Err(anyhow!("empty turn samples"));
+    }
+    let samples_16k: Vec<i16> = turn.samples.iter().step_by(3).copied().collect();
+    if samples_16k.is_empty() {
+        return Err(anyhow!("empty 16k asr samples"));
+    }
+    let samples_per_chunk = 320usize;
+    Ok(samples_16k
+        .chunks(samples_per_chunk)
+        .map(|chunk| chunk.to_vec())
+        .collect())
+}
+
+fn pcm_i16_to_le_bytes(samples: &[i16]) -> Vec<u8> {
+    let mut bytes = Vec::with_capacity(samples.len() * 2);
+    for sample in samples {
+        bytes.extend_from_slice(&sample.to_le_bytes());
+    }
+    bytes
+}
+
 async fn request_turn_bridge_stream(
     http: &Client,
     bridge_config: &TurnBridgeConfig,
@@ -1013,6 +1374,7 @@
     turn: &FinishedSpeechTurn,
     path_ref: &str,
     byte_size: u64,
+    asr_result_ref: Option<&str>,
     turn_pipeline_started_at: Instant,
 ) -> Result<RuntimeTurnStreamOutcome> {
     let bridge_started_at = Instant::now();
@@ -1020,7 +1382,7 @@
         call_id: call_id.to_string(),
         trace_id: trace_id.to_string(),
         turn_id: turn.turn_id.clone(),
-        audio_artifact: RuntimeTurnAudioArtifact {
+        audio_artifact: Some(RuntimeTurnAudioArtifact {
             artifact_type: "local_file".to_string(),
             path_ref: path_ref.to_string(),
             format: "wav".to_string(),
@@ -1028,7 +1390,8 @@
             channels: u32::from(USER_AUDIO_NUM_CHANNELS),
             duration_ms: turn.duration_ms,
             byte_size,
-        },
+        }),
+        asr_result_ref: asr_result_ref.map(str::to_string),
     };
     let response = http
         .post(bridge_config.bridge_url.as_deref().unwrap_or_default())
@@ -1107,6 +1470,7 @@
             .await?;
         }
     }
+    state.close_timing();
     if !state.completed {
         warn!(
             call_id = %call_id,
@@ -1236,6 +1600,7 @@
                 call_id,
                 trace_id,
                 turn,
+                &event,
                 audio_chunk,
                 state,
                 turn_pipeline_started_at,
@@ -1247,6 +1612,9 @@
         }
         Some("device_output") => {
             if let Some(output) = event.device_output.as_ref() {
+                if !should_publish_device_output(&mut state.published_device_output_ids, output) {
+                    return Ok(());
+                }
                 sink.publish_device_output(call_id, trace_id, &turn.turn_id, output)
                     .await?;
                 state.device_output_count = state.device_output_count.saturating_add(1);
@@ -1254,6 +1622,7 @@
         }
         Some("turn_completed") => {
             state.completed = true;
+            state.close_timing();
             info!(
                 call_id = %call_id,
                 trace_id = %trace_id,
@@ -1264,6 +1633,7 @@
             );
         }
         Some("turn_failed") => {
+            state.close_timing();
             let error = event.error.as_ref();
             let reason_code = error
                 .and_then(|value| value.reason_code.as_deref())
@@ -1298,6 +1668,7 @@
         }
         Some("turn_cancelled") => {
             state.completed = true;
+            state.close_timing();
             info!(
                 call_id = %call_id,
                 trace_id = %trace_id,
@@ -1330,6 +1701,19 @@
         }
         Some("activity") => {
             if let Some(activity) = event.activity.as_ref() {
+                if activity.event_type.as_deref() == Some("tts_first_audio_chunk_ready") {
+                    if let Some(runtime_session_nonce) =
+                        bridge_config.runtime_session_nonce.as_deref()
+                    {
+                        let _ = state.arm_timing_anchor(
+                            call_id,
+                            trace_id,
+                            &turn.turn_id,
+                            runtime_session_nonce,
+                            &event,
+                        );
+                    }
+                }
                 info!(
                     call_id = %call_id,
                     trace_id = %trace_id,
@@ -1388,6 +1772,7 @@
     call_id: &str,
     trace_id: &str,
     turn: &FinishedSpeechTurn,
+    event: &RuntimeTurnStreamEvent,
     audio_chunk: &RuntimeTurnStreamAudioChunk,
     state: &mut RuntimeTurnStreamState,
     turn_pipeline_started_at: Instant,
@@ -1407,6 +1792,58 @@
         .unwrap_or("pcm_s16le")
         .trim()
         .to_ascii_lowercase();
+    if !matches!(format.as_str(), "pcm_s16le" | "mp3" | "mpeg" | "wav") {
+        return Err(anyhow!("unsupported reply_audio_chunk format {format}"));
+    }
+    match state.reply_chunk_markers.observe(audio_chunk.segment_seq) {
+        ReplyChunkMarker::FirstReply => {
+            let extension = json!({
+                "segmentSeq": audio_chunk.segment_seq,
+                "chunkSeq": audio_chunk.chunk_seq,
+                "format": format.as_str(),
+                "bytes": payload.len(),
+            });
+            if let Some(marker) =
+                state.record_m6(call_id, trace_id, &turn.turn_id, event, audio_chunk)
+            {
+                emit_anchored_activity(
+                    call_id,
+                    trace_id,
+                    &turn.turn_id,
+                    "helper_first_reply_audio_chunk_received",
+                    &marker,
+                    extension,
+                );
+            } else {
+                emit_activity(
+                    call_id,
+                    trace_id,
+                    Some(&turn.turn_id),
+                    "helper_first_reply_audio_chunk_received",
+                    "ok",
+                    None,
+                    None,
+                    extension,
+                );
+            }
+        }
+        ReplyChunkMarker::SegmentFirst => emit_activity(
+            call_id,
+            trace_id,
+            Some(&turn.turn_id),
+            "helper_segment_first_audio_chunk_received",
+            "ok",
+            None,
+            None,
+            json!({
+                "segmentSeq": audio_chunk.segment_seq,
+                "chunkSeq": audio_chunk.chunk_seq,
+                "format": format.as_str(),
+                "bytes": payload.len(),
+            }),
+        ),
+        ReplyChunkMarker::None => {}
+    }
     let frames = if format == "pcm_s16le" {
         let sample_rate = audio_chunk.sample_rate.unwrap_or(sink.sample_rate_hz);
         let channels = audio_chunk.channels.unwrap_or(u32::from(sink.num_channels));
@@ -1498,7 +1935,7 @@
             Err(error) => return Err(error).context("failed to decode final stream audio chunk"),
         }
     } else {
-        return Err(anyhow!("unsupported reply_audio_chunk format {format}"));
+        unreachable!("supported encoded format checked above")
     };
     if frames.is_empty() {
         return Ok(0);
@@ -1526,21 +1963,33 @@
         sink.write_pcm_frame(frame).await?;
         if !state.first_audio_frame_written {
             state.first_audio_frame_written = true;
-            emit_activity(
-                call_id,
-                trace_id,
-                Some(&turn.turn_id),
-                "bot_reply_first_audio_frame_written",
-                "ok",
-                None,
-                None,
-                json!({
-                    "replyPlaybackMode": state.reply_playback_mode.as_str(),
-                    "format": format.as_str(),
-                    "chunkSeq": audio_chunk.chunk_seq,
-                    "replyTotalAfterVadEndMs": turn_pipeline_started_at.elapsed().as_millis() as u64,
-                }),
-            );
+            let extension = json!({
+                "replyPlaybackMode": state.reply_playback_mode.as_str(),
+                "format": format.as_str(),
+                "chunkSeq": audio_chunk.chunk_seq,
+                "replyTotalAfterVadEndMs": turn_pipeline_started_at.elapsed().as_millis() as u64,
+            });
+            if let Some(marker) = state.record_m7() {
+                emit_anchored_activity(
+                    call_id,
+                    trace_id,
+                    &turn.turn_id,
+                    "bot_reply_first_audio_frame_written",
+                    &marker,
+                    extension,
+                );
+            } else {
+                emit_activity(
+                    call_id,
+                    trace_id,
+                    Some(&turn.turn_id),
+                    "bot_reply_first_audio_frame_written",
+                    "ok",
+                    None,
+                    None,
+                    extension,
+                );
+            }
         }
         sleep_until(pacing_started_at + Duration::from_millis(((index + 1) as u64) * 20)).await;
     }
@@ -1621,6 +2070,7 @@
     turn: &FinishedSpeechTurn,
     path_ref: &str,
     byte_size: u64,
+    asr_result_ref: Option<&str>,
     turn_pipeline_started_at: Instant,
 ) -> Result<RuntimeTurnBridgeOutcome> {
     let bridge_started_at = Instant::now();
@@ -1628,7 +2078,7 @@
         call_id: call_id.to_string(),
         trace_id: trace_id.to_string(),
         turn_id: turn.turn_id.clone(),
-        audio_artifact: RuntimeTurnAudioArtifact {
+        audio_artifact: Some(RuntimeTurnAudioArtifact {
             artifact_type: "local_file".to_string(),
             path_ref: path_ref.to_string(),
             format: "wav".to_string(),
@@ -1636,7 +2086,8 @@
             channels: u32::from(USER_AUDIO_NUM_CHANNELS),
             duration_ms: turn.duration_ms,
             byte_size,
-        },
+        }),
+        asr_result_ref: asr_result_ref.map(str::to_string),
     };
     let response = http
         .post(bridge_config.bridge_url.as_deref().unwrap_or_default())
@@ -1927,7 +2378,10 @@
     #[serde(rename = "turnId")]
     turn_id: String,
     #[serde(rename = "audioArtifact")]
-    audio_artifact: RuntimeTurnAudioArtifact,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    audio_artifact: Option<RuntimeTurnAudioArtifact>,
+    #[serde(rename = "asrResultRef", skip_serializing_if = "Option::is_none")]
+    asr_result_ref: Option<String>,
 }
 
 #[derive(Serialize)]
@@ -1955,6 +2409,20 @@
     retryable: Option<bool>,
 }
 
+#[derive(Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct RuntimeAsrStreamResp {
+    status: Option<String>,
+    asr_result_ref: Option<String>,
+    chunk_count: Option<u64>,
+    audio_bytes: Option<u64>,
+    asr_duration_ms: Option<u64>,
+    provider_alias: Option<String>,
+    text_len: Option<u64>,
+    fallback_reason: Option<String>,
+    fallback_stage: Option<String>,
+}
+
 #[derive(Default)]
 struct RuntimeTurnBridgeOutcome {
     reply_audio_artifact: Option<RuntimeTurnReplyAudioArtifact>,
@@ -1975,9 +2443,12 @@
     completed: bool,
     audio_chunk_count: u64,
     device_output_count: u64,
+    published_device_output_ids: HashSet<String>,
     encoded_audio_buffer: Vec<u8>,
     pcm_stream_decoder: Option<audio::PcmS16leStreamDecoder>,
     pcm_stream_network_chunk_count: u64,
+    reply_chunk_markers: ReplyChunkMarkerState,
+    timing: RuntimeTurnStreamTimingState,
 }
 
 impl Default for RuntimeTurnStreamState {
@@ -1990,10 +2461,266 @@
             completed: false,
             audio_chunk_count: 0,
             device_output_count: 0,
+            published_device_output_ids: HashSet::new(),
             encoded_audio_buffer: Vec::new(),
             pcm_stream_decoder: None,
             pcm_stream_network_chunk_count: 0,
+            reply_chunk_markers: ReplyChunkMarkerState::default(),
+            timing: RuntimeTurnStreamTimingState::default(),
         }
+    }
+}
+
+impl RuntimeTurnStreamState {
+    fn arm_timing_anchor(
+        &mut self,
+        call_id: &str,
+        trace_id: &str,
+        turn_id: &str,
+        runtime_session_nonce: &str,
+        event: &RuntimeTurnStreamEvent,
+    ) -> bool {
+        if self.timing.phase != RuntimeTurnStreamTimingPhase::Empty
+            || event.call_id.as_deref() != Some(call_id)
+            || event.trace_id.as_deref() != Some(trace_id)
+            || event.turn_id.as_deref() != Some(turn_id)
+        {
+            return false;
+        }
+        let Some(activity) = event.activity.as_ref() else {
+            return false;
+        };
+        if activity.event_type.as_deref() != Some("tts_first_audio_chunk_ready") {
+            return false;
+        }
+        let Some(extension) = activity.extension.as_ref() else {
+            return false;
+        };
+        let Some(anchor_id) = extension.stream_anchor_id.as_deref() else {
+            return false;
+        };
+        let valid_anchor_id = (16..=64).contains(&anchor_id.len()) && anchor_id.is_ascii();
+        let expected_nonce_hash = runtime_session_nonce_hash(runtime_session_nonce);
+        if extension.stream_timing_version != Some(STREAM_TIMING_VERSION)
+            || !valid_anchor_id
+            || extension.runtime_session_nonce_hash.as_deref() != Some(expected_nonce_hash.as_str())
+            || extension.segment_seq != Some(STREAM_TIMING_FIRST_SEGMENT)
+            || extension.stream_timing_validation.as_deref() != Some("bound")
+        {
+            return false;
+        }
+        let Some(server_delta_ms) = extension.stream_anchor_server_delta_ms else {
+            return false;
+        };
+        self.timing.anchor = Some(RuntimeTurnStreamTimingAnchor {
+            call_id: call_id.to_string(),
+            trace_id: trace_id.to_string(),
+            turn_id: turn_id.to_string(),
+            anchor_id: anchor_id.to_string(),
+            server_delta_ms,
+            runtime_session_nonce_hash: expected_nonce_hash,
+            segment_seq: STREAM_TIMING_FIRST_SEGMENT,
+            received_at: Instant::now(),
+        });
+        self.timing.phase = RuntimeTurnStreamTimingPhase::Armed;
+        true
+    }
+
+    fn record_m6(
+        &mut self,
+        call_id: &str,
+        trace_id: &str,
+        turn_id: &str,
+        event: &RuntimeTurnStreamEvent,
+        audio_chunk: &RuntimeTurnStreamAudioChunk,
+    ) -> Option<RuntimeTurnStreamTimingMarker> {
+        if self.timing.phase != RuntimeTurnStreamTimingPhase::Armed {
+            return None;
+        }
+        let anchor = self.timing.anchor.as_ref()?;
+        if anchor.call_id != call_id
+            || anchor.trace_id != trace_id
+            || anchor.turn_id != turn_id
+            || event.call_id.as_deref() != Some(call_id)
+            || event.trace_id.as_deref() != Some(trace_id)
+            || event.turn_id.as_deref() != Some(turn_id)
+            || audio_chunk.segment_seq != Some(anchor.segment_seq)
+            || audio_chunk.chunk_seq != Some(STREAM_TIMING_FIRST_CHUNK)
+            || audio_chunk.stream_timing_version != Some(STREAM_TIMING_VERSION)
+            || audio_chunk.stream_anchor_id.as_deref() != Some(anchor.anchor_id.as_str())
+        {
+            return None;
+        }
+        let Some(marker) = RuntimeTurnStreamTimingMarker::from_anchor(
+            anchor,
+            audio_chunk.chunk_seq.unwrap_or(STREAM_TIMING_FIRST_CHUNK),
+        ) else {
+            self.close_timing();
+            return None;
+        };
+        self.timing.chunk_seq = Some(marker.chunk_seq);
+        self.timing.phase = RuntimeTurnStreamTimingPhase::M6Recorded;
+        Some(marker)
+    }
+
+    fn record_m7(&mut self) -> Option<RuntimeTurnStreamTimingMarker> {
+        if self.timing.phase != RuntimeTurnStreamTimingPhase::M6Recorded {
+            return None;
+        }
+        let anchor = self.timing.anchor.as_ref()?;
+        let Some(marker) = RuntimeTurnStreamTimingMarker::from_anchor(
+            anchor,
+            self.timing.chunk_seq.unwrap_or(STREAM_TIMING_FIRST_CHUNK),
+        ) else {
+            self.close_timing();
+            return None;
+        };
+        self.timing.phase = RuntimeTurnStreamTimingPhase::M7Recorded;
+        Some(marker)
+    }
+
+    fn close_timing(&mut self) {
+        self.timing.close();
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum RuntimeTurnStreamTimingPhase {
+    Empty,
+    Armed,
+    M6Recorded,
+    M7Recorded,
+    Closed,
+}
+
+struct RuntimeTurnStreamTimingState {
+    phase: RuntimeTurnStreamTimingPhase,
+    anchor: Option<RuntimeTurnStreamTimingAnchor>,
+    chunk_seq: Option<u64>,
+}
+
+impl Default for RuntimeTurnStreamTimingState {
+    fn default() -> Self {
+        Self {
+            phase: RuntimeTurnStreamTimingPhase::Empty,
+            anchor: None,
+            chunk_seq: None,
+        }
+    }
+}
+
+impl RuntimeTurnStreamTimingState {
+    fn close(&mut self) {
+        self.anchor = None;
+        self.chunk_seq = None;
+        self.phase = RuntimeTurnStreamTimingPhase::Closed;
+    }
+}
+
+impl Drop for RuntimeTurnStreamTimingState {
+    fn drop(&mut self) {
+        self.anchor = None;
+        self.chunk_seq = None;
+    }
+}
+
+struct RuntimeTurnStreamTimingAnchor {
+    call_id: String,
+    trace_id: String,
+    turn_id: String,
+    anchor_id: String,
+    server_delta_ms: u64,
+    runtime_session_nonce_hash: String,
+    segment_seq: u64,
+    received_at: Instant,
+}
+
+struct RuntimeTurnStreamTimingMarker {
+    version: u32,
+    anchor_id: String,
+    anchor_server_delta_ms: u64,
+    anchor_elapsed_ms: u64,
+    server_delta_ms: u64,
+    runtime_session_nonce_hash: String,
+    segment_seq: u64,
+    chunk_seq: u64,
+}
+
+impl RuntimeTurnStreamTimingMarker {
+    fn from_anchor(anchor: &RuntimeTurnStreamTimingAnchor, chunk_seq: u64) -> Option<Self> {
+        let elapsed_ms = u64::try_from(anchor.received_at.elapsed().as_millis()).ok()?;
+        if elapsed_ms > STREAM_TIMING_MAX_ELAPSED_MS {
+            return None;
+        }
+        Some(Self {
+            version: STREAM_TIMING_VERSION,
+            anchor_id: anchor.anchor_id.clone(),
+            anchor_server_delta_ms: anchor.server_delta_ms,
+            anchor_elapsed_ms: elapsed_ms,
+            server_delta_ms: anchor.server_delta_ms.checked_add(elapsed_ms)?,
+            runtime_session_nonce_hash: anchor.runtime_session_nonce_hash.clone(),
+            segment_seq: anchor.segment_seq,
+            chunk_seq,
+        })
+    }
+
+    fn extension_with(&self, extra: serde_json::Value) -> serde_json::Value {
+        let mut extension = match extra {
+            serde_json::Value::Object(value) => value,
+            _ => serde_json::Map::new(),
+        };
+        extension.insert("streamTimingVersion".to_string(), json!(self.version));
+        extension.insert("streamAnchorId".to_string(), json!(self.anchor_id));
+        extension.insert(
+            "streamAnchorServerDeltaMs".to_string(),
+            json!(self.anchor_server_delta_ms),
+        );
+        extension.insert("anchorElapsedMs".to_string(), json!(self.anchor_elapsed_ms));
+        extension.insert(
+            "runtimeSessionNonceHash".to_string(),
+            json!(self.runtime_session_nonce_hash),
+        );
+        extension.insert("segmentSeq".to_string(), json!(self.segment_seq));
+        extension.insert("chunkSeq".to_string(), json!(self.chunk_seq));
+        extension.insert("streamTimingValidation".to_string(), json!("bound"));
+        serde_json::Value::Object(extension)
+    }
+}
+
+fn runtime_session_nonce_hash(value: &str) -> String {
+    let digest = Sha256::digest(value.as_bytes());
+    digest[..6]
+        .iter()
+        .map(|byte| format!("{byte:02x}"))
+        .collect()
+}
+
+#[derive(Debug, PartialEq, Eq)]
+enum ReplyChunkMarker {
+    FirstReply,
+    SegmentFirst,
+    None,
+}
+
+#[derive(Default)]
+struct ReplyChunkMarkerState {
+    first_reply_seen: bool,
+    seen_segments: HashSet<u64>,
+}
+
+impl ReplyChunkMarkerState {
+    fn observe(&mut self, segment_seq: Option<u64>) -> ReplyChunkMarker {
+        let first_for_segment = segment_seq
+            .map(|value| self.seen_segments.insert(value))
+            .unwrap_or(false);
+        if !self.first_reply_seen {
+            self.first_reply_seen = true;
+            return ReplyChunkMarker::FirstReply;
+        }
+        if first_for_segment {
+            return ReplyChunkMarker::SegmentFirst;
+        }
+        ReplyChunkMarker::None
     }
 }
 
@@ -2013,6 +2740,12 @@
 struct RuntimeTurnStreamEvent {
     #[serde(rename = "type", alias = "event")]
     event_type: Option<String>,
+    #[serde(rename = "callId")]
+    call_id: Option<String>,
+    #[serde(rename = "traceId")]
+    trace_id: Option<String>,
+    #[serde(rename = "turnId")]
+    turn_id: Option<String>,
     #[serde(rename = "seq")]
     seq: Option<u64>,
     #[serde(rename = "replyPlaybackMode")]
@@ -2048,6 +2781,12 @@
 struct RuntimeTurnStreamAudioChunk {
     #[serde(rename = "chunkSeq", alias = "seq")]
     chunk_seq: Option<u64>,
+    #[serde(rename = "segmentSeq")]
+    segment_seq: Option<u64>,
+    #[serde(rename = "streamTimingVersion")]
+    stream_timing_version: Option<u32>,
+    #[serde(rename = "streamAnchorId")]
+    stream_anchor_id: Option<String>,
     format: Option<String>,
     #[serde(rename = "sampleRate")]
     sample_rate: Option<u32>,
@@ -2064,6 +2803,23 @@
     stage: Option<String>,
     #[serde(rename = "reasonCode")]
     reason_code: Option<String>,
+    extension: Option<RuntimeTurnStreamTimingExtension>,
+}
+
+#[derive(Clone, Deserialize)]
+struct RuntimeTurnStreamTimingExtension {
+    #[serde(rename = "streamTimingVersion")]
+    stream_timing_version: Option<u32>,
+    #[serde(rename = "streamAnchorId")]
+    stream_anchor_id: Option<String>,
+    #[serde(rename = "streamAnchorServerDeltaMs")]
+    stream_anchor_server_delta_ms: Option<u64>,
+    #[serde(rename = "runtimeSessionNonceHash")]
+    runtime_session_nonce_hash: Option<String>,
+    #[serde(rename = "segmentSeq")]
+    segment_seq: Option<u64>,
+    #[serde(rename = "streamTimingValidation")]
+    stream_timing_validation: Option<String>,
 }
 
 #[derive(Deserialize)]
@@ -2096,6 +2852,30 @@
     #[serde(rename = "commandCode")]
     command_code: Option<String>,
     params: Option<serde_json::Value>,
+}
+
+fn should_publish_device_output(
+    published_ids: &mut HashSet<String>,
+    output: &RuntimeTurnDeviceOutput,
+) -> bool {
+    let Some(command_id) = output
+        .command_id
+        .as_deref()
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+    else {
+        return false;
+    };
+    if output
+        .command_code
+        .as_deref()
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+        .is_none()
+    {
+        return false;
+    }
+    published_ids.insert(command_id.to_string())
 }
 
 fn require_safe_segment(value: &str) -> Result<()> {
@@ -2137,6 +2917,7 @@
     turn_bridge_config: TurnBridgeConfig,
     http: Client,
     sink: Arc<BotAudioOutputSink>,
+    participant: RemoteParticipant,
 ) -> JoinHandle<()> {
     tokio::spawn(async move {
         let mut stream = NativeAudioStream::new(
@@ -2145,6 +2926,68 @@
             i32::from(USER_AUDIO_NUM_CHANNELS),
         );
         let started_at = Instant::now();
+        let (frame_tx, mut frame_rx) =
+            mpsc::channel::<DrainedUserAudioFrame>(INBOUND_AUDIO_QUEUE_CAPACITY);
+        let (drain_shutdown_tx, mut drain_shutdown_rx) = watch::channel(false);
+        let drain_call_id = call_id.clone();
+        let drain_trace_id = trace_id.clone();
+        let drain_task = tokio::spawn(async move {
+            let mut received_frame_count: u64 = 0;
+            let mut dropped_frame_count: u64 = 0;
+            loop {
+                tokio::select! {
+                    changed = drain_shutdown_rx.changed() => {
+                        match changed {
+                            Ok(()) if *drain_shutdown_rx.borrow() => break,
+                            Ok(()) => {}
+                            Err(_) => break,
+                        }
+                    }
+                    maybe_frame = stream.next() => {
+                        let Some(frame) = maybe_frame else {
+                            break;
+                        };
+                        received_frame_count = received_frame_count.saturating_add(1);
+                        let drained = DrainedUserAudioFrame {
+                            frame_index: received_frame_count,
+                            captured_elapsed_ms: started_at.elapsed().as_millis() as u64,
+                            frame: AudioFrame {
+                                data: Cow::Owned(frame.data.as_ref().to_vec()),
+                                sample_rate: frame.sample_rate,
+                                num_channels: frame.num_channels,
+                                samples_per_channel: frame.samples_per_channel,
+                            },
+                        };
+                        match frame_tx.try_send(drained) {
+                            Ok(()) => {}
+                            Err(mpsc::error::TrySendError::Full(_)) => {
+                                dropped_frame_count = dropped_frame_count.saturating_add(1);
+                                if dropped_frame_count % INBOUND_AUDIO_DROP_LOG_INTERVAL == 1 {
+                                    warn!(
+                                        call_id = %drain_call_id,
+                                        trace_id = %drain_trace_id,
+                                        dropped_frame_count,
+                                        received_frame_count,
+                                        queue_capacity = INBOUND_AUDIO_QUEUE_CAPACITY,
+                                        "runtime helper inbound_audio_queue_full_dropping_newest"
+                                    );
+                                }
+                            }
+                            Err(mpsc::error::TrySendError::Closed(_)) => break,
+                        }
+                    }
+                }
+            }
+            stream.close();
+            info!(
+                call_id = %drain_call_id,
+                trace_id = %drain_trace_id,
+                received_frame_count,
+                dropped_frame_count,
+                queue_capacity = INBOUND_AUDIO_QUEUE_CAPACITY,
+                "runtime helper user_audio_drain_ended"
+            );
+        });
         let mut frame_count: u64 = 0;
         let mut sample_count: u64 = 0;
         let mut simple_vad = if simple_vad_enabled {
@@ -2152,11 +2995,13 @@
         } else {
             None
         };
+        let mut realtime_asr_upload: Option<RealtimeAsrUpload> = None;
 
-        while let Some(frame) = stream.next().await {
-            frame_count += 1;
+        while let Some(drained) = frame_rx.recv().await {
+            let frame = drained.frame;
+            frame_count = drained.frame_index;
             sample_count += u64::from(frame.samples_per_channel) * u64::from(frame.num_channels);
-            let elapsed_ms = started_at.elapsed().as_millis() as u64;
+            let elapsed_ms = drained.captured_elapsed_ms;
             if frame_count == 1 {
                 info!(
                     call_id = %call_id,
@@ -2184,7 +3029,8 @@
 
             if let Some(vad) = simple_vad.as_mut() {
                 if vad_enabled_gate.load(Ordering::Acquire) {
-                    if let Some(turn) = vad.observe_frame(
+                    let was_in_speech = vad.in_speech;
+                    let turn = vad.observe_frame(
                         &call_id,
                         &trace_id,
                         &participant_alias,
@@ -2192,7 +3038,66 @@
                         frame_count,
                         elapsed_ms,
                         &frame,
-                    ) {
+                    );
+                    let is_in_speech = vad.in_speech;
+
+                    if !was_in_speech && is_in_speech {
+                        let turn_id = format!("turn-{:04}", vad.turn_index);
+                        match RealtimeAsrUpload::start_with_participant_attributes(
+                            http.clone(),
+                            turn_bridge_config.realtime_asr_config(),
+                            &call_id,
+                            &trace_id,
+                            &turn_id,
+                            &vad.speech_samples,
+                            || participant.attributes(),
+                        ) {
+                            Ok(upload) => {
+                                info!(
+                                    call_id = %call_id,
+                                    trace_id = %trace_id,
+                                    turn_id = %turn_id,
+                                    "runtime helper asr_realtime_session_started"
+                                );
+                                realtime_asr_upload = Some(upload);
+                            }
+                            Err(error) if turn_bridge_config.asr_realtime_enabled => {
+                                warn!(
+                                    call_id = %call_id,
+                                    trace_id = %trace_id,
+                                    turn_id = %turn_id,
+                                    error = %safe_error(&error.to_string()),
+                                    "runtime helper asr_realtime_start_failed_fallback"
+                                );
+                            }
+                            Err(_) => {}
+                        }
+                    } else if was_in_speech {
+                        let push_failed = realtime_asr_upload
+                            .as_mut()
+                            .and_then(|upload| upload.push_48k_samples(frame.data.as_ref()).err());
+                        if let Some(error) = push_failed {
+                            warn!(
+                                call_id = %call_id,
+                                trace_id = %trace_id,
+                                error = %safe_error(&error.to_string()),
+                                "runtime helper asr_realtime_upload_failed_fallback"
+                            );
+                            if let Some(upload) = realtime_asr_upload.take() {
+                                tokio::spawn(async move {
+                                    upload.cancel("upload_backpressure").await;
+                                });
+                            }
+                        }
+                    }
+
+                    if let Some(turn) = turn {
+                        let realtime_asr_result_ref = match realtime_asr_upload.take() {
+                            Some(upload) => {
+                                finish_realtime_asr_upload(upload, &call_id, &trace_id, &turn).await
+                            }
+                            None => None,
+                        };
                         handle_finished_turn(
                             &http,
                             &turn_bridge_config,
@@ -2200,10 +3105,22 @@
                             &call_id,
                             &trace_id,
                             turn,
+                            realtime_asr_result_ref,
                         )
                         .await;
+                    } else if was_in_speech && !is_in_speech {
+                        if let Some(upload) = realtime_asr_upload.take() {
+                            tokio::spawn(async move {
+                                upload.cancel("speech_too_short").await;
+                            });
+                        }
                     }
                 } else {
+                    if let Some(upload) = realtime_asr_upload.take() {
+                        tokio::spawn(async move {
+                            upload.cancel("vad_disabled").await;
+                        });
+                    }
                     vad.observe_disabled_frame(
                         &call_id,
                         &trace_id,
@@ -2224,9 +3141,41 @@
                 &track_sid_alias,
                 started_at.elapsed().as_millis() as u64,
             ) {
-                handle_finished_turn(&http, &turn_bridge_config, &sink, &call_id, &trace_id, turn)
-                    .await;
+                let realtime_asr_result_ref = match realtime_asr_upload.take() {
+                    Some(upload) => {
+                        finish_realtime_asr_upload(upload, &call_id, &trace_id, &turn).await
+                    }
+                    None => None,
+                };
+                handle_finished_turn(
+                    &http,
+                    &turn_bridge_config,
+                    &sink,
+                    &call_id,
+                    &trace_id,
+                    turn,
+                    realtime_asr_result_ref,
+                )
+                .await;
             }
+        }
+        if let Some(upload) = realtime_asr_upload.take() {
+            upload.cancel("stream_end").await;
+        }
+        let _ = drain_shutdown_tx.send(true);
+        let mut drain_task = drain_task;
+        if timeout(AUDIO_DRAIN_STOP_GRACE, &mut drain_task)
+            .await
+            .is_err()
+        {
+            drain_task.abort();
+            let _ = drain_task.await;
+            warn!(
+                call_id = %call_id,
+                trace_id = %trace_id,
+                stop_grace_ms = AUDIO_DRAIN_STOP_GRACE.as_millis() as u64,
+                "runtime helper user_audio_drain_stop_timeout"
+            );
         }
 
         info!(
@@ -2240,6 +3189,64 @@
             "runtime helper user_audio_stream_ended"
         );
     })
+}
+
+struct DrainedUserAudioFrame {
+    frame_index: u64,
+    captured_elapsed_ms: u64,
+    frame: AudioFrame<'static>,
+}
+
+async fn finish_realtime_asr_upload(
+    upload: RealtimeAsrUpload,
+    call_id: &str,
+    trace_id: &str,
+    turn: &FinishedSpeechTurn,
+) -> Option<String> {
+    match upload.finish(turn.duration_ms, &turn.end_reason).await {
+        Ok(RealtimeAsrOutcome {
+            status,
+            asr_result_ref,
+            provider_alias,
+            partial_count,
+            fallback_reason,
+            fallback_stage,
+            chunk_count,
+            audio_bytes,
+            wall_ms,
+        }) => {
+            info!(
+                call_id = %call_id,
+                trace_id = %trace_id,
+                turn_id = %turn.turn_id,
+                status = %status,
+                provider_alias = ?provider_alias,
+                partial_count,
+                chunk_count,
+                audio_bytes,
+                wall_ms,
+                asr_result_ref_present = asr_result_ref.is_some(),
+                fallback_reason = ?fallback_reason,
+                fallback_stage = ?fallback_stage,
+                "runtime helper asr_realtime_finished"
+            );
+            if status == "final" {
+                asr_result_ref
+            } else {
+                None
+            }
+        }
+        Err(error) => {
+            warn!(
+                call_id = %call_id,
+                trace_id = %trace_id,
+                turn_id = %turn.turn_id,
+                error = %safe_error(&error.to_string()),
+                "runtime helper asr_realtime_failed_fallback"
+            );
+            None
+        }
+    }
 }
 
 #[derive(Clone)]
@@ -2933,3 +3940,142 @@
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::{
+        ReplyChunkMarker, ReplyChunkMarkerState, RuntimeTurnDeviceOutput, RuntimeTurnStreamEvent,
+        RuntimeTurnStreamState, RuntimeTurnStreamTimingPhase, runtime_session_nonce_hash,
+        should_publish_device_output,
+    };
+    use std::collections::HashSet;
+
+    #[test]
+    fn reply_chunk_marker_state_emits_turn_first_once_and_later_segment_first_once() {
+        let mut state = ReplyChunkMarkerState::default();
+
+        assert_eq!(ReplyChunkMarker::FirstReply, state.observe(Some(1)));
+        assert_eq!(ReplyChunkMarker::None, state.observe(Some(1)));
+        assert_eq!(ReplyChunkMarker::SegmentFirst, state.observe(Some(2)));
+        assert_eq!(ReplyChunkMarker::None, state.observe(Some(2)));
+        assert_eq!(ReplyChunkMarker::SegmentFirst, state.observe(Some(3)));
+    }
+
+    #[test]
+    fn reply_chunk_marker_state_without_segment_only_emits_turn_first() {
+        let mut state = ReplyChunkMarkerState::default();
+
+        assert_eq!(ReplyChunkMarker::FirstReply, state.observe(None));
+        assert_eq!(ReplyChunkMarker::None, state.observe(None));
+    }
+
+    #[test]
+    fn runtime_turn_stream_audio_chunk_reads_segment_seq() {
+        let event: RuntimeTurnStreamEvent = serde_json::from_str(
+            r#"{"type":"reply_audio_chunk","audioChunk":{"chunkSeq":4,"segmentSeq":2,"format":"pcm_s16le","payloadBase64":"AA==","last":false}}"#,
+        )
+        .expect("turn stream event");
+
+        assert_eq!(
+            Some(2),
+            event.audio_chunk.and_then(|chunk| chunk.segment_seq)
+        );
+    }
+
+    #[test]
+    fn runtime_turn_stream_reads_frozen_m5_and_audio_chunk_timing_contract() {
+        let activity: RuntimeTurnStreamEvent = serde_json::from_str(
+            r#"{"type":"activity","callId":"call-1","traceId":"trace-1","turnId":"turn-1","activity":{"eventType":"tts_first_audio_chunk_ready","extension":{"streamTimingVersion":1,"streamAnchorId":"0123456789abcdef","streamAnchorServerDeltaMs":1200,"runtimeSessionNonceHash":"abcdef012345","segmentSeq":1,"streamTimingValidation":"bound"}}}"#,
+        )
+        .expect("m5 activity event");
+        let audio: RuntimeTurnStreamEvent = serde_json::from_str(
+            r#"{"type":"reply_audio_chunk","callId":"call-1","traceId":"trace-1","turnId":"turn-1","audioChunk":{"chunkSeq":1,"segmentSeq":1,"streamTimingVersion":1,"streamAnchorId":"0123456789abcdef","format":"pcm_s16le","payloadBase64":"AA==","last":false}}"#,
+        )
+        .expect("timed audio chunk event");
+
+        assert_eq!(Some("call-1"), activity.call_id.as_deref());
+        assert_eq!(Some("trace-1"), activity.trace_id.as_deref());
+        assert_eq!(Some("turn-1"), activity.turn_id.as_deref());
+        let extension = activity.activity.unwrap().extension.unwrap();
+        assert_eq!(Some(1), extension.stream_timing_version);
+        assert_eq!(Some(1200), extension.stream_anchor_server_delta_ms);
+        assert_eq!(
+            Some("0123456789abcdef"),
+            audio.audio_chunk.unwrap().stream_anchor_id.as_deref()
+        );
+    }
+
+    #[test]
+    fn runtime_turn_stream_timing_records_m6_and_m7_once_then_rejects_terminal_late_events() {
+        let nonce = "runtime-nonce";
+        let nonce_hash = runtime_session_nonce_hash(nonce);
+        let activity: RuntimeTurnStreamEvent = serde_json::from_str(&format!(
+            r#"{{"type":"activity","callId":"call-1","traceId":"trace-1","turnId":"turn-1","activity":{{"eventType":"tts_first_audio_chunk_ready","extension":{{"streamTimingVersion":1,"streamAnchorId":"0123456789abcdef","streamAnchorServerDeltaMs":1200,"runtimeSessionNonceHash":"{nonce_hash}","segmentSeq":1,"streamTimingValidation":"bound"}}}}}}"#,
+        ))
+        .expect("m5 activity event");
+        let audio: RuntimeTurnStreamEvent = serde_json::from_str(
+            r#"{"type":"reply_audio_chunk","callId":"call-1","traceId":"trace-1","turnId":"turn-1","audioChunk":{"chunkSeq":1,"segmentSeq":1,"streamTimingVersion":1,"streamAnchorId":"0123456789abcdef","format":"pcm_s16le","payloadBase64":"AA==","last":false}}"#,
+        )
+        .expect("timed audio chunk event");
+        let mut state = RuntimeTurnStreamState::default();
+
+        assert!(state.arm_timing_anchor("call-1", "trace-1", "turn-1", nonce, &activity));
+        let chunk = audio.audio_chunk.as_ref().expect("audio chunk");
+        assert!(
+            state
+                .record_m6("call-1", "trace-1", "turn-1", &audio, chunk)
+                .is_some()
+        );
+        assert!(
+            state
+                .record_m6("call-1", "trace-1", "turn-1", &audio, chunk)
+                .is_none()
+        );
+        assert!(state.record_m7().is_some());
+        assert!(state.record_m7().is_none());
+
+        state.close_timing();
+        assert_eq!(RuntimeTurnStreamTimingPhase::Closed, state.timing.phase);
+        assert!(state.timing.anchor.is_none());
+        assert!(!state.arm_timing_anchor("call-1", "trace-1", "turn-1", nonce, &activity));
+        assert!(
+            state
+                .record_m6("call-1", "trace-1", "turn-1", &audio, chunk)
+                .is_none()
+        );
+        assert!(state.record_m7().is_none());
+    }
+
+    #[test]
+    fn device_output_contract_is_reliable_and_deduplicated() {
+        let output = RuntimeTurnDeviceOutput {
+            command_id: Some("cmd-1".to_string()),
+            command_code: Some("custom.app.DeviceLevelChange".to_string()),
+            params: Some(serde_json::json!({"level": 1})),
+        };
+        let mut published = HashSet::new();
+        assert!(should_publish_device_output(&mut published, &output));
+        assert!(!should_publish_device_output(&mut published, &output));
+        assert_eq!(published.len(), 1);
+    }
+
+    #[test]
+    fn device_output_invalid_or_missing_command_is_fail_closed() {
+        for output in [
+            RuntimeTurnDeviceOutput {
+                command_id: None,
+                command_code: Some("custom.app.DeviceLevelChange".to_string()),
+                params: None,
+            },
+            RuntimeTurnDeviceOutput {
+                command_id: Some("cmd-1".to_string()),
+                command_code: None,
+                params: None,
+            },
+        ] {
+            let mut published = HashSet::new();
+            assert!(!should_publish_device_output(&mut published, &output));
+            assert!(published.is_empty());
+        }
+    }
+}

--
Gitblit v1.9.3