//go:build darwin
|
|
package engine
|
|
import (
|
"fmt"
|
"voicesnap/internal/logger"
|
"voicesnap/internal/model"
|
|
_ "github.com/k2-fsa/sherpa-onnx-go-macos"
|
sherpa "github.com/k2-fsa/sherpa-onnx-go/sherpa_onnx"
|
)
|
|
type sherpaEngine struct {
|
recognizer *sherpa.OfflineRecognizer
|
hwInfo string
|
}
|
|
type xasrStreamingEngine struct {
|
recognizer *sherpa.OnlineRecognizer
|
hwInfo string
|
}
|
|
func newPlatformEngine(resolved model.ResolvedModel) (Engine, error) {
|
if !isSupportedBackend(resolved.BackendKind) {
|
return nil, fmt.Errorf("unsupported backend kind: %s", resolved.BackendKind)
|
}
|
if resolved.BackendKind == model.BackendXASRStreaming {
|
return newXASRStreamingEngine(resolved)
|
}
|
|
providers := darwinProviders(resolved.ProviderOrder)
|
|
for _, p := range providers {
|
config, err := offlineConfigForResolvedModel(resolved, p.provider)
|
if err != nil {
|
return nil, err
|
}
|
|
recognizer := sherpa.NewOfflineRecognizer(&config)
|
if recognizer != nil {
|
info := fmt.Sprintf("%s · %s", resolved.Profile.DisplayName, p.name)
|
logger.Info("Engine initialized: %s", info)
|
return &sherpaEngine{
|
recognizer: recognizer,
|
hwInfo: info,
|
}, nil
|
}
|
logger.Info("Failed to init with %s, trying next provider", p.name)
|
}
|
|
return nil, fmt.Errorf("failed to initialize sherpa-onnx with any provider")
|
}
|
|
func newXASRStreamingEngine(resolved model.ResolvedModel) (Engine, error) {
|
providers := darwinProviders(resolved.ProviderOrder)
|
|
for _, p := range providers {
|
config, err := onlineConfigForResolvedModel(resolved, p.provider)
|
if err != nil {
|
return nil, err
|
}
|
|
recognizer := sherpa.NewOnlineRecognizer(&config)
|
if recognizer != nil {
|
info := fmt.Sprintf("%s · %s", resolved.Profile.DisplayName, p.name)
|
logger.Info("Engine initialized: %s", info)
|
return &xasrStreamingEngine{
|
recognizer: recognizer,
|
hwInfo: info,
|
}, nil
|
}
|
logger.Info("Failed to init X-ASR with %s, trying next provider", p.name)
|
}
|
|
return nil, fmt.Errorf("failed to initialize X-ASR streaming recognizer with any provider")
|
}
|
|
func offlineConfigForResolvedModel(resolved model.ResolvedModel, provider string) (sherpa.OfflineRecognizerConfig, error) {
|
config := sherpa.OfflineRecognizerConfig{}
|
config.FeatConfig.SampleRate = 16000
|
config.FeatConfig.FeatureDim = 80
|
config.ModelConfig.Tokens = resolved.Files["tokens"]
|
config.ModelConfig.NumThreads = resolved.Profile.NumThreads
|
config.ModelConfig.Provider = provider
|
config.DecodingMethod = "greedy_search"
|
|
switch resolved.BackendKind {
|
case model.BackendSenseVoice:
|
config.ModelConfig.SenseVoice.Model = resolved.Files["model"]
|
config.ModelConfig.SenseVoice.UseInverseTextNormalization = 1
|
case model.BackendMoonshine:
|
config.ModelConfig.Moonshine.Preprocessor = resolved.Files["preprocessor"]
|
config.ModelConfig.Moonshine.Encoder = resolved.Files["encoder"]
|
config.ModelConfig.Moonshine.UncachedDecoder = resolved.Files["uncached_decoder"]
|
config.ModelConfig.Moonshine.CachedDecoder = resolved.Files["cached_decoder"]
|
case model.BackendNemoTransducer:
|
config.ModelConfig.Transducer.Encoder = resolved.Files["encoder"]
|
config.ModelConfig.Transducer.Decoder = resolved.Files["decoder"]
|
config.ModelConfig.Transducer.Joiner = resolved.Files["joiner"]
|
config.ModelConfig.ModelType = model.BackendNemoTransducer
|
case model.BackendQwen3ASR:
|
config.ModelConfig.Qwen3ASR.ConvFrontend = resolved.Files["conv_frontend"]
|
config.ModelConfig.Qwen3ASR.Encoder = resolved.Files["encoder"]
|
config.ModelConfig.Qwen3ASR.Decoder = resolved.Files["decoder"]
|
config.ModelConfig.Qwen3ASR.Tokenizer = resolved.Files["tokenizer"]
|
config.ModelConfig.Qwen3ASR.MaxTotalLen = 1024
|
config.ModelConfig.Qwen3ASR.MaxNewTokens = 256
|
config.ModelConfig.Qwen3ASR.Temperature = 0.0
|
config.ModelConfig.Qwen3ASR.TopP = 0.9
|
config.ModelConfig.Qwen3ASR.Seed = 0
|
default:
|
return sherpa.OfflineRecognizerConfig{}, fmt.Errorf("unsupported backend kind: %s", resolved.BackendKind)
|
}
|
|
return config, nil
|
}
|
|
func onlineConfigForResolvedModel(resolved model.ResolvedModel, provider string) (sherpa.OnlineRecognizerConfig, error) {
|
config := sherpa.OnlineRecognizerConfig{}
|
config.FeatConfig.SampleRate = 16000
|
config.FeatConfig.FeatureDim = 80
|
config.ModelConfig.Tokens = resolved.Files["tokens"]
|
config.ModelConfig.NumThreads = resolved.Profile.NumThreads
|
config.ModelConfig.Provider = provider
|
config.ModelConfig.ModelType = "zipformer2"
|
config.DecodingMethod = "greedy_search"
|
config.EnableEndpoint = 0
|
|
switch resolved.BackendKind {
|
case model.BackendXASRStreaming:
|
config.ModelConfig.Transducer.Encoder = resolved.Files["encoder"]
|
config.ModelConfig.Transducer.Decoder = resolved.Files["decoder"]
|
config.ModelConfig.Transducer.Joiner = resolved.Files["joiner"]
|
default:
|
return sherpa.OnlineRecognizerConfig{}, fmt.Errorf("unsupported online backend kind: %s", resolved.BackendKind)
|
}
|
|
return config, nil
|
}
|
|
func darwinProviders(providerOrder []string) []struct {
|
name string
|
provider string
|
} {
|
names := map[string]string{
|
"coreml": "CoreML (Apple Neural Engine)",
|
"cpu": "CPU",
|
}
|
providers := make([]struct {
|
name string
|
provider string
|
}, 0, len(providerOrder))
|
for _, provider := range providerOrder {
|
name, ok := names[provider]
|
if !ok {
|
continue
|
}
|
providers = append(providers, struct {
|
name string
|
provider string
|
}{name: name, provider: provider})
|
}
|
if len(providers) == 0 {
|
providers = append(providers, struct {
|
name string
|
provider string
|
}{name: "CPU", provider: "cpu"})
|
}
|
return providers
|
}
|
|
func (e *sherpaEngine) Recognize(samples []float32) (string, error) {
|
stream := sherpa.NewOfflineStream(e.recognizer)
|
defer sherpa.DeleteOfflineStream(stream)
|
|
stream.AcceptWaveform(16000, samples)
|
|
e.recognizer.Decode(stream)
|
result := stream.GetResult()
|
|
return result.Text, nil
|
}
|
|
func (e *xasrStreamingEngine) Recognize(samples []float32) (string, error) {
|
stream := sherpa.NewOnlineStream(e.recognizer)
|
defer sherpa.DeleteOnlineStream(stream)
|
|
if len(samples) > 0 {
|
stream.AcceptWaveform(16000, samples)
|
}
|
stream.InputFinished()
|
|
for e.recognizer.IsReady(stream) {
|
e.recognizer.Decode(stream)
|
}
|
|
result := e.recognizer.GetResult(stream)
|
if result == nil {
|
return "", nil
|
}
|
|
return result.Text, nil
|
}
|
|
func (e *sherpaEngine) HardwareInfo() string {
|
return e.hwInfo
|
}
|
|
func (e *xasrStreamingEngine) HardwareInfo() string {
|
return e.hwInfo
|
}
|
|
func (e *sherpaEngine) Close() {
|
if e.recognizer != nil {
|
sherpa.DeleteOfflineRecognizer(e.recognizer)
|
e.recognizer = nil
|
}
|
}
|
|
func (e *xasrStreamingEngine) Close() {
|
if e.recognizer != nil {
|
sherpa.DeleteOnlineRecognizer(e.recognizer)
|
e.recognizer = nil
|
}
|
}
|
