From 0ffbf1935c9d091cce22a5583275ac0902f7a693 Mon Sep 17 00:00:00 2001
From: Ariver <shanghai3168@gmail.com>
Date: Thu, 02 Jul 2026 03:04:51 +0800
Subject: [PATCH] Add Cantonese language model support
---
privatevoice.src/internal/engine/engine_darwin_test.go | 231 ++++++++++++++++++++++++++++++++++++++++++++++++++++-----
1 files changed, 208 insertions(+), 23 deletions(-)
diff --git a/privatevoice.src/internal/engine/engine_darwin_test.go b/privatevoice.src/internal/engine/engine_darwin_test.go
index 8310a03..55f5f0a 100644
--- a/privatevoice.src/internal/engine/engine_darwin_test.go
+++ b/privatevoice.src/internal/engine/engine_darwin_test.go
@@ -3,8 +3,10 @@
package engine
import (
+ "os"
"path/filepath"
"testing"
+ "time"
"voicesnap/internal/model"
)
@@ -50,20 +52,60 @@
}
}
-func TestOfflineConfigForQwen3ASRUsesQwen3Config(t *testing.T) {
+func TestOfflineConfigForZipformerUsesPlainTransducer(t *testing.T) {
root := t.TempDir()
files := map[string]string{
- "conv_frontend": filepath.Join(root, "conv_frontend.onnx"),
- "encoder": filepath.Join(root, "encoder.int8.onnx"),
- "decoder": filepath.Join(root, "decoder.int8.onnx"),
- "tokenizer": filepath.Join(root, "tokenizer"),
+ "encoder": filepath.Join(root, "encoder-epoch-99-avg-1.int8.onnx"),
+ "decoder": filepath.Join(root, "decoder-epoch-99-avg-1.onnx"),
+ "joiner": filepath.Join(root, "joiner-epoch-99-avg-1.int8.onnx"),
+ "tokens": filepath.Join(root, "tokens.txt"),
}
resolved := model.ResolvedModel{
- ModelID: model.Qwen3ASRModelID,
- BackendKind: model.BackendQwen3ASR,
+ ModelID: model.JapaneseModelID,
+ BackendKind: model.BackendTransducer,
Profile: model.ModelProfile{
- DisplayName: "Qwen3-ASR",
- BackendKind: model.BackendQwen3ASR,
+ DisplayName: "Zipformer Japanese",
+ BackendKind: model.BackendTransducer,
+ ProviderOrder: []string{"cpu"},
+ NumThreads: 1,
+ },
+ Files: files,
+ }
+
+ config, err := offlineConfigForResolvedModel(resolved, "cpu")
+ if err != nil {
+ t.Fatal(err)
+ }
+ if config.ModelConfig.Transducer.Encoder != files["encoder"] {
+ t.Fatalf("encoder = %q, want %q", config.ModelConfig.Transducer.Encoder, files["encoder"])
+ }
+ if config.ModelConfig.Transducer.Decoder != files["decoder"] {
+ t.Fatalf("decoder = %q, want %q", config.ModelConfig.Transducer.Decoder, files["decoder"])
+ }
+ if config.ModelConfig.Transducer.Joiner != files["joiner"] {
+ t.Fatalf("joiner = %q, want %q", config.ModelConfig.Transducer.Joiner, files["joiner"])
+ }
+ if config.ModelConfig.Tokens != files["tokens"] {
+ t.Fatalf("tokens = %q, want %q", config.ModelConfig.Tokens, files["tokens"])
+ }
+ if config.ModelConfig.ModelType != "" {
+ t.Fatalf("model type = %q, want empty for plain transducer", config.ModelConfig.ModelType)
+ }
+}
+
+func TestOfflineConfigForSenseVoiceYueSetsLanguageParam(t *testing.T) {
+ root := t.TempDir()
+ files := map[string]string{
+ "model": filepath.Join(root, "model.int8.onnx"),
+ "tokens": filepath.Join(root, "tokens.txt"),
+ }
+ resolved := model.ResolvedModel{
+ ModelID: model.SenseVoiceYueModelID,
+ BackendKind: model.BackendSenseVoice,
+ Profile: model.ModelProfile{
+ DisplayName: "SenseVoice Yue",
+ BackendKind: model.BackendSenseVoice,
+ LanguageParam: "yue",
ProviderOrder: []string{"cpu"},
NumThreads: 4,
},
@@ -74,25 +116,168 @@
if err != nil {
t.Fatal(err)
}
- if config.ModelConfig.Qwen3ASR.ConvFrontend != files["conv_frontend"] {
- t.Fatalf("conv frontend = %q, want %q", config.ModelConfig.Qwen3ASR.ConvFrontend, files["conv_frontend"])
+ if config.ModelConfig.SenseVoice.Model != files["model"] {
+ t.Fatalf("model = %q, want %q", config.ModelConfig.SenseVoice.Model, files["model"])
}
- if config.ModelConfig.Qwen3ASR.Encoder != files["encoder"] {
- t.Fatalf("encoder = %q, want %q", config.ModelConfig.Qwen3ASR.Encoder, files["encoder"])
+ if config.ModelConfig.SenseVoice.Language != "yue" {
+ t.Fatalf("language = %q, want yue", config.ModelConfig.SenseVoice.Language)
}
- if config.ModelConfig.Qwen3ASR.Decoder != files["decoder"] {
- t.Fatalf("decoder = %q, want %q", config.ModelConfig.Qwen3ASR.Decoder, files["decoder"])
+ if config.ModelConfig.Tokens != files["tokens"] {
+ t.Fatalf("tokens = %q, want %q", config.ModelConfig.Tokens, files["tokens"])
}
- if config.ModelConfig.Qwen3ASR.Tokenizer != files["tokenizer"] {
- t.Fatalf("tokenizer = %q, want %q", config.ModelConfig.Qwen3ASR.Tokenizer, files["tokenizer"])
+}
+
+func TestOnlineConfigForXASRUsesStreamingTransducer(t *testing.T) {
+ root := t.TempDir()
+ files := map[string]string{
+ "encoder": filepath.Join(root, "encoder-960ms.onnx"),
+ "decoder": filepath.Join(root, "decoder-960ms.onnx"),
+ "joiner": filepath.Join(root, "joiner-960ms.onnx"),
+ "tokens": filepath.Join(root, "tokens.txt"),
}
- if config.ModelConfig.Qwen3ASR.MaxTotalLen != 1024 {
- t.Fatalf("max total len = %d, want 1024", config.ModelConfig.Qwen3ASR.MaxTotalLen)
+ resolved := model.ResolvedModel{
+ ModelID: model.XASRZhEn960ModelID,
+ BackendKind: model.BackendXASRStreaming,
+ Profile: model.ModelProfile{
+ DisplayName: "X-ASR zh-en 960ms",
+ BackendKind: model.BackendXASRStreaming,
+ ProviderOrder: []string{"cpu"},
+ NumThreads: 1,
+ },
+ Files: files,
}
- if config.ModelConfig.Qwen3ASR.MaxNewTokens != 256 {
- t.Fatalf("max new tokens = %d, want 256", config.ModelConfig.Qwen3ASR.MaxNewTokens)
+
+ config, err := onlineConfigForResolvedModel(resolved, "cpu")
+ if err != nil {
+ t.Fatal(err)
}
- if config.ModelConfig.Tokens != "" {
- t.Fatalf("tokens = %q, want empty for qwen3-asr", config.ModelConfig.Tokens)
+ if config.ModelConfig.Transducer.Encoder != files["encoder"] {
+ t.Fatalf("encoder = %q, want %q", config.ModelConfig.Transducer.Encoder, files["encoder"])
}
+ if config.ModelConfig.Transducer.Decoder != files["decoder"] {
+ t.Fatalf("decoder = %q, want %q", config.ModelConfig.Transducer.Decoder, files["decoder"])
+ }
+ if config.ModelConfig.Transducer.Joiner != files["joiner"] {
+ t.Fatalf("joiner = %q, want %q", config.ModelConfig.Transducer.Joiner, files["joiner"])
+ }
+ if config.ModelConfig.Tokens != files["tokens"] {
+ t.Fatalf("tokens = %q, want %q", config.ModelConfig.Tokens, files["tokens"])
+ }
+ if config.ModelConfig.ModelType != "zipformer2" {
+ t.Fatalf("model type = %q, want zipformer2", config.ModelConfig.ModelType)
+ }
+ if config.DecodingMethod != "greedy_search" {
+ t.Fatalf("decoding method = %q, want greedy_search", config.DecodingMethod)
+ }
+}
+
+func TestRememberNonEmptyTextKeepsLastPartial(t *testing.T) {
+ got := rememberNonEmptyText("最后一句", " ")
+ if got != "最后一句" {
+ t.Fatalf("rememberNonEmptyText() = %q, want last non-empty text", got)
+ }
+
+ got = rememberNonEmptyText("旧内容", "新内容")
+ if got != "新内容" {
+ t.Fatalf("rememberNonEmptyText() = %q, want new non-empty text", got)
+ }
+}
+
+func TestXASRTailPaddingMatchesFlushWindow(t *testing.T) {
+ if xasrTailPaddingSamples != 24000 {
+ t.Fatalf("tail padding samples = %d, want 24000", xasrTailPaddingSamples)
+ }
+ if len(xasrTailPadding) != xasrTailPaddingSamples {
+ t.Fatalf("tail padding len = %d, want %d", len(xasrTailPadding), xasrTailPaddingSamples)
+ }
+}
+
+func TestXASRHeadPaddingPreparesInitialContext(t *testing.T) {
+ if xasrHeadPaddingSamples != 4000 {
+ t.Fatalf("head padding samples = %d, want 4000", xasrHeadPaddingSamples)
+ }
+ input := []float32{0.1, -0.2, 0.3}
+ got := xasrSamplesWithHeadPadding(input)
+
+ if len(got) != xasrHeadPaddingSamples+len(input) {
+ t.Fatalf("padded len = %d, want %d", len(got), xasrHeadPaddingSamples+len(input))
+ }
+ for i := 0; i < xasrHeadPaddingSamples; i++ {
+ if got[i] != 0 {
+ t.Fatalf("head padding sample %d = %v, want 0", i, got[i])
+ }
+ }
+ for i, want := range input {
+ if got[xasrHeadPaddingSamples+i] != want {
+ t.Fatalf("payload sample %d = %v, want %v", i, got[xasrHeadPaddingSamples+i], want)
+ }
+ }
+}
+
+func TestXASRRequestsReleaseTailCaptureDelay(t *testing.T) {
+ got := (&xasrStreamingEngine{}).ReleaseTailCaptureDelay()
+ if got != 300*time.Millisecond {
+ t.Fatalf("ReleaseTailCaptureDelay() = %v, want 300ms", got)
+ }
+}
+
+func TestSenseVoiceRequestsReleaseTailCaptureDelay(t *testing.T) {
+ got := (&sherpaEngine{backendKind: model.BackendSenseVoice}).ReleaseTailCaptureDelay()
+ if got != 100*time.Millisecond {
+ t.Fatalf("ReleaseTailCaptureDelay() = %v, want 100ms", got)
+ }
+}
+
+func TestOtherSherpaModelsDoNotRequestReleaseTailCaptureDelay(t *testing.T) {
+ for _, backendKind := range []string{
+ model.BackendMoonshine,
+ model.BackendTransducer,
+ model.BackendNemoTransducer,
+ model.BackendQwen3ASR,
+ } {
+ t.Run(backendKind, func(t *testing.T) {
+ got := (&sherpaEngine{backendKind: backendKind}).ReleaseTailCaptureDelay()
+ if got != 0 {
+ t.Fatalf("ReleaseTailCaptureDelay() = %v, want 0", got)
+ }
+ })
+ }
+}
+
+func TestXASREnablesHoldPreCapture(t *testing.T) {
+ if !(&xasrStreamingEngine{}).HoldPreCaptureEnabled() {
+ t.Fatal("X-ASR should enable hold pre-capture")
+ }
+}
+
+func TestOfflineSherpaDoesNotEnableHoldPreCapture(t *testing.T) {
+ if _, ok := any(&sherpaEngine{}).(HoldPreCaptureEngine); ok {
+ t.Fatal("offline sherpa engine should not enable hold pre-capture")
+ }
+}
+
+func TestXASRRealModelSmoke(t *testing.T) {
+ if os.Getenv("PRIVATEVOICE_XASR_SMOKE") != "1" {
+ t.Skip("set PRIVATEVOICE_XASR_SMOKE=1 to run the real X-ASR model smoke test")
+ }
+
+ resolved, err := model.ResolveModel(model.XASRZhEn960ModelID)
+ if err != nil {
+ t.Fatal(err)
+ }
+ if !resolved.IsUsable() {
+ t.Fatalf("x-asr model is not usable: status=%s missing=%v problems=%v", resolved.Status, resolved.Missing, resolved.Problems)
+ }
+
+ eng, err := NewWithResolvedModel(resolved)
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer eng.Close()
+
+ text, err := eng.Recognize(make([]float32, 16000))
+ if err != nil {
+ t.Fatal(err)
+ }
+ t.Logf("hardware=%s silence_text=%q", eng.HardwareInfo(), text)
}
--
Gitblit v1.9.3