mirror of
https://github.com/mudler/LocalAI.git
synced 2026-10-05 04:24:39 -04:00
* feat(voice): list registered voices and record which encoder made them The voice registry could register, identify and forget but not list, and it did not remember which speaker encoder produced an embedding. Add Metadata.Model and Registry.List, answered from the index the store registry already keeps for Forget. Needed so a backend can be given the registered voices that match its own speaker encoder. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * feat(voice): store the encoder model with a registered voice Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * feat(voice): pick the registered voices that match a speaker model Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * feat(proto): carry known voices and speaker names on diarize and live messages Assisted-by: Claude:claude-haiku-4-5 [Claude Code] * feat(diarization): name speakers from the voice registry When a diarization model has a speaker_model option, the endpoint sends the registered voices made by that encoder to the backend. The backend's name and name_score come back as extra fields next to the normalized SPEAKER_NN speaker, and the speakers summary carries the first name seen for each speaker. RTTM output and results without names are unchanged. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * feat(live): pass registered voices to a live session and surface speaker names Live sessions now send the registered voices that match the model's speaker_model to the backend, and each speaker segment carries the name the backend matched. The realtime segment event gains an optional speaker_name field. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * feat(parakeet-cpp): load a speaker model and build per-request voice registries Adds the speaker bindings (ABI v9 and v10, probed separately), the speaker_model, speaker_threshold and speaker_margin options, and a per-request registry builder over the known voices. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * feat(parakeet-cpp): name the speakers in Diarize from the known voices Diarize builds a per-request speaker registry from the known voices when a speaker model is loaded, calls the named C functions, and puts each slot's registered name and score on the segments. The registry is freed on every path. A library without ABI 10 reports Unimplemented instead of dropping the names. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * feat(parakeet-cpp): name speakers in the live scene stream The live scene stream now begins with a known-voice registry when a speaker model is loaded and the live config carries voices, and each closed speaker segment takes its slot's current name from the feed's names map. A segment that closes before its slot is identified has an empty name. The registry is freed after the stream, on every path. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * feat(gallery): speaker naming entries and docs for parakeet-cpp Add three gallery entries that load the WeSpeaker ResNet34 speaker model next to the diarization or realtime scene models, and document speaker names in the voice recognition, diarization, audio to text and realtime pages. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * fix(parakeet-cpp): skip an unusable registered voice instead of failing the request A registered voice with the wrong embedding size, or one the C side refused, failed the whole diarization request, so one legacy voice broke the model for every user. Skip such voices with a warning that does not carry the voice name, and take the plain path when none is left. Also map an exact 0 speaker threshold or margin to a tiny positive value, since the C side reads 0 as "use the default", and fix a stale comment about which contexts Free() walks. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * fix(diarization): warn once per model about voices from another encoder; document the privacy limit The different-encoder warning fired on every request. Log it once per feature and speaker model, then at debug level. Document that the global voice registry lets any caller of a speaker_model model learn matching names, and that skipped wrong-sized voices are logged. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] * chore(parakeet-cpp): bump parakeet.cpp to 8c8cec0 (C-API v10) and check speaker naming against the real library The pin moves from 623a968 to 8c8cec0, which brings in everything merged in parakeet.cpp since: the voice identification change (C-API v9, #78) and raw-embedding enroll plus diarize-only speaker naming (C-API v10, #79). New real-library specs (gated on PARAKEET_BACKEND_TEST_SPEAKER_MODEL, _DIAR_MODEL, _WAV and, for the live path, _STREAM_MODEL) name the two speakers of two_speakers.wav from a committed pair of WeSpeaker embeddings, with the voices passed in reversed order. They also check that the float32 threshold reaches C through purego. The shared test loader now registers the v9/v10 and scene symbols as main.go does. The rebase onto origin/master had no conflicts. Assisted-by: Claude:claude-sonnet-5-5 [Claude Code] --------- Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
379 lines
11 KiB
Go
379 lines
11 KiB
Go
package backend
|
|
|
|
import (
|
|
"context"
|
|
"errors"
|
|
"fmt"
|
|
"io"
|
|
"maps"
|
|
"sync"
|
|
"time"
|
|
|
|
"github.com/mudler/LocalAI/core/config"
|
|
"github.com/mudler/LocalAI/core/schema"
|
|
"github.com/mudler/LocalAI/core/services/voicerecognition"
|
|
"github.com/mudler/LocalAI/core/trace"
|
|
grpcPkg "github.com/mudler/LocalAI/pkg/grpc"
|
|
"github.com/mudler/LocalAI/pkg/grpc/proto"
|
|
"github.com/mudler/LocalAI/pkg/model"
|
|
"github.com/mudler/LocalAI/pkg/sound"
|
|
"github.com/mudler/xlog"
|
|
)
|
|
|
|
// LiveTranscriptionEvent is one streamed event from a live (bidirectional)
|
|
// transcription session. Delta/Eou/Eob/Words arrive as the user speaks; Final
|
|
// is set exactly once, on the terminal event after Close flushes the decode
|
|
// tail. Eou means the model judged the user yielded the turn; Eob means a
|
|
// backchannel ("uh-huh") ended — callers must NOT treat Eob as a turn
|
|
// boundary.
|
|
type LiveTranscriptionEvent struct {
|
|
Delta string
|
|
Eou bool
|
|
Eob bool
|
|
Words []schema.TranscriptionWord
|
|
Speakers []LiveSpeakerSegment
|
|
Sounds []LiveSoundEvent
|
|
Final *schema.TranscriptionResult
|
|
}
|
|
|
|
// LiveSpeakerSegment is one closed speaker segment from a companion
|
|
// diarization/scene stream running alongside live transcription. Start/End
|
|
// are stream-relative seconds (mapped from the backend's nanoseconds).
|
|
type LiveSpeakerSegment struct {
|
|
Speaker string
|
|
// Name is the registered speaker name the backend matched, empty when the
|
|
// speaker is unknown.
|
|
Name string
|
|
Start float64
|
|
End float64
|
|
}
|
|
|
|
// LiveSoundEvent is one closed sound event from a companion sound/scene
|
|
// stream running alongside live transcription. Start/End are stream-relative
|
|
// seconds (mapped from the backend's nanoseconds).
|
|
type LiveSoundEvent struct {
|
|
Label string
|
|
Index int
|
|
Peak float32
|
|
Start float64
|
|
End float64
|
|
}
|
|
|
|
// LiveTranscriptionSession is a handle on an open live transcription stream.
|
|
// Feed pushes 16 kHz mono float PCM; Close signals end-of-audio, waits for
|
|
// the backend's terminal Final event to be delivered, and releases the
|
|
// stream.
|
|
type LiveTranscriptionSession interface {
|
|
Feed(pcm []float32) error
|
|
Close() error
|
|
}
|
|
|
|
// liveCloseDrainTimeout bounds how long Close waits for the backend to flush
|
|
// the decode tail before force-cancelling the stream. Finalize is one short
|
|
// engine call; seconds here means the backend is wedged.
|
|
const liveCloseDrainTimeout = 10 * time.Second
|
|
|
|
type liveTranscriptionSession struct {
|
|
stream grpcPkg.AudioTranscriptionLiveClient
|
|
cancel context.CancelFunc
|
|
recvDone chan struct{}
|
|
recvErr error // written by the recv goroutine before recvDone closes
|
|
closeOnce sync.Once
|
|
closeErr error
|
|
trace *liveTraceState // nil when tracing was disabled at open
|
|
release func()
|
|
}
|
|
|
|
func (s *liveTranscriptionSession) Feed(pcm []float32) error {
|
|
s.trace.addPCM(pcm)
|
|
return s.stream.Send(&proto.TranscriptLiveRequest{
|
|
Payload: &proto.TranscriptLiveRequest_Audio{Audio: &proto.TranscriptLiveAudio{Pcm: pcm}},
|
|
})
|
|
}
|
|
|
|
func (s *liveTranscriptionSession) Close() error {
|
|
s.closeOnce.Do(func() {
|
|
err := s.stream.CloseSend()
|
|
select {
|
|
case <-s.recvDone:
|
|
case <-time.After(liveCloseDrainTimeout):
|
|
xlog.Warn("live transcription: backend did not finalize in time; cancelling stream")
|
|
s.cancel()
|
|
<-s.recvDone
|
|
}
|
|
s.cancel()
|
|
if err == nil {
|
|
err = s.recvErr
|
|
}
|
|
s.closeErr = err
|
|
s.trace.record(err)
|
|
s.release()
|
|
})
|
|
return s.closeErr
|
|
}
|
|
|
|
// liveSampleRate is the PCM rate of a live transcription session, fixed by
|
|
// the session config sent in ModelTranscriptionLive.
|
|
const liveSampleRate = 16000
|
|
|
|
// liveTraceState accumulates what the per-turn backend trace needs while a
|
|
// live session runs: a bounded copy of the fed PCM for the audio snippet,
|
|
// the decode outputs, and timing. One trace is recorded at Close — the live
|
|
// path never touches the unary transcription wrapper, so without this a
|
|
// streaming-only pipeline produced no transcription traces at all. Feed and
|
|
// the recv goroutine run concurrently; mu guards the accumulators.
|
|
type liveTraceState struct {
|
|
appConfig *config.ApplicationConfig
|
|
modelName string
|
|
backend string
|
|
language string
|
|
started time.Time
|
|
traceID string
|
|
|
|
mu sync.Mutex
|
|
pcm []byte // first trace.MaxSnippetSeconds of fed audio, int16 LE
|
|
fedSamples int // ALL samples fed, beyond the snippet cap
|
|
deltaEvents int
|
|
eouEvents int
|
|
eobEvents int
|
|
finalText string
|
|
}
|
|
|
|
func newLiveTraceState(modelConfig config.ModelConfig, appConfig *config.ApplicationConfig, language string) *liveTraceState {
|
|
if !appConfig.EnableTracing {
|
|
return nil
|
|
}
|
|
trace.InitBackendTracingIfEnabled(appConfig.TracingMaxItems, appConfig.TracingMaxBodyBytes)
|
|
started := time.Now()
|
|
return &liveTraceState{
|
|
appConfig: appConfig,
|
|
modelName: modelConfig.Name,
|
|
backend: modelConfig.Backend,
|
|
language: language,
|
|
started: started,
|
|
traceID: trace.BeginBackendTrace(trace.BackendTrace{Timestamp: started, Type: trace.BackendTraceTranscription, ModelName: modelConfig.Name, Backend: modelConfig.Backend, Summary: "live transcription"}),
|
|
}
|
|
}
|
|
|
|
func (ts *liveTraceState) addPCM(pcm []float32) {
|
|
if ts == nil {
|
|
return
|
|
}
|
|
ts.mu.Lock()
|
|
defer ts.mu.Unlock()
|
|
ts.fedSamples += len(pcm)
|
|
maxBytes := trace.MaxSnippetSeconds * liveSampleRate * 2
|
|
if room := (maxBytes - len(ts.pcm)) / 2; room > 0 {
|
|
if len(pcm) > room {
|
|
pcm = pcm[:room]
|
|
}
|
|
ts.pcm = append(ts.pcm, sound.Float32sToInt16LEBytes(pcm)...)
|
|
}
|
|
}
|
|
|
|
func (ts *liveTraceState) observe(ev LiveTranscriptionEvent) {
|
|
if ts == nil {
|
|
return
|
|
}
|
|
ts.mu.Lock()
|
|
defer ts.mu.Unlock()
|
|
if ev.Delta != "" {
|
|
ts.deltaEvents++
|
|
}
|
|
if ev.Eou {
|
|
ts.eouEvents++
|
|
}
|
|
if ev.Eob {
|
|
ts.eobEvents++
|
|
}
|
|
if ev.Final != nil {
|
|
ts.finalText = ev.Final.Text
|
|
}
|
|
}
|
|
|
|
func (ts *liveTraceState) record(closeErr error) {
|
|
if ts == nil || !ts.appConfig.EnableTracing {
|
|
return
|
|
}
|
|
ts.mu.Lock()
|
|
data := map[string]any{
|
|
"source": "live_stream",
|
|
"language": ts.language,
|
|
"result_text": ts.finalText,
|
|
"eou_events": ts.eouEvents,
|
|
"eob_events": ts.eobEvents,
|
|
"delta_events": ts.deltaEvents,
|
|
}
|
|
if snippet := trace.AudioSnippetFromPCM(ts.pcm, liveSampleRate, ts.fedSamples*2, ts.appConfig.TracingMaxBodyBytes); snippet != nil {
|
|
maps.Copy(data, snippet)
|
|
}
|
|
summary := "live -> " + ts.finalText
|
|
ts.mu.Unlock()
|
|
|
|
bt := trace.BackendTrace{
|
|
ID: ts.traceID,
|
|
Timestamp: ts.started,
|
|
Duration: time.Since(ts.started),
|
|
Type: trace.BackendTraceTranscription,
|
|
ModelName: ts.modelName,
|
|
Backend: ts.backend,
|
|
Summary: trace.TruncateString(summary, 200),
|
|
Data: data,
|
|
}
|
|
if closeErr != nil {
|
|
bt.Error = closeErr.Error()
|
|
}
|
|
trace.RecordBackendTrace(bt)
|
|
}
|
|
|
|
// LiveOption tunes a live transcription session.
|
|
type LiveOption func(*liveOptions)
|
|
|
|
type liveOptions struct {
|
|
knownVoices []voicerecognition.KnownVoice
|
|
}
|
|
|
|
// WithKnownVoices gives the backend the registered voices it may use to name
|
|
// the speakers it detects. Backends without speaker identification ignore them.
|
|
func WithKnownVoices(v []voicerecognition.KnownVoice) LiveOption {
|
|
return func(o *liveOptions) { o.knownVoices = v }
|
|
}
|
|
|
|
// liveConfigProto builds the first message of a live session.
|
|
func liveConfigProto(language string, o liveOptions) *proto.TranscriptLiveConfig {
|
|
cfg := &proto.TranscriptLiveConfig{Language: language, SampleRate: liveSampleRate}
|
|
for _, v := range o.knownVoices {
|
|
cfg.KnownVoices = append(cfg.KnownVoices, &proto.KnownVoice{Name: v.Name, Embedding: v.Embedding, Model: v.Model})
|
|
}
|
|
return cfg
|
|
}
|
|
|
|
// ModelTranscriptionLive loads the transcription backend, opens the
|
|
// bidirectional AudioTranscriptionLive RPC, sends the session config, and
|
|
// BLOCKS until the backend's ready ack. A grpcerrors.
|
|
// IsLiveTranscriptionUnsupported error means the backend (or the loaded
|
|
// model) cannot do live transcription and the caller should degrade to the
|
|
// unary/file path. After a successful return, onEvent is invoked from a
|
|
// background goroutine — in order, one event at a time — for every response
|
|
// the backend streams, ending with the Final event triggered by Close.
|
|
func ModelTranscriptionLive(ctx context.Context, language string,
|
|
ml *model.ModelLoader, modelConfig config.ModelConfig, appConfig *config.ApplicationConfig,
|
|
onEvent func(LiveTranscriptionEvent), opts ...LiveOption) (LiveTranscriptionSession, error) {
|
|
|
|
lo := liveOptions{}
|
|
for _, f := range opts {
|
|
f(&lo)
|
|
}
|
|
|
|
transcriptionModel, err := loadTranscriptionModel(ctx, ml, modelConfig, appConfig)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
release, err := AcquireGlobalBackendSlot()
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
|
|
// The derived cancel out-lives this call inside the session: Close uses
|
|
// it to unwind the stream (and, in embed mode, the server-side recv
|
|
// pump, which only stops on send-close or context cancellation).
|
|
streamCtx, cancel := context.WithCancel(ctx)
|
|
stream, err := transcriptionModel.AudioTranscriptionLive(streamCtx)
|
|
if err != nil {
|
|
cancel()
|
|
release()
|
|
return nil, err
|
|
}
|
|
|
|
fail := func(err error) (LiveTranscriptionSession, error) {
|
|
_ = stream.CloseSend()
|
|
cancel()
|
|
release()
|
|
return nil, err
|
|
}
|
|
|
|
if err := stream.Send(&proto.TranscriptLiveRequest{
|
|
Payload: &proto.TranscriptLiveRequest_Config{Config: liveConfigProto(language, lo)},
|
|
}); err != nil {
|
|
return fail(err)
|
|
}
|
|
|
|
// Ready-ack contract: the backend answers a successful open with a
|
|
// {ready:true} response before any transcript data; unsupported
|
|
// backends surface Unimplemented here instead.
|
|
ack, err := stream.Recv()
|
|
if err != nil {
|
|
return fail(err)
|
|
}
|
|
if !ack.GetReady() {
|
|
return fail(fmt.Errorf("live transcription: backend %q broke the ready-ack contract (first response carried data)", modelConfig.Backend))
|
|
}
|
|
|
|
s := &liveTranscriptionSession{
|
|
stream: stream,
|
|
cancel: cancel,
|
|
recvDone: make(chan struct{}),
|
|
trace: newLiveTraceState(modelConfig, appConfig, language),
|
|
release: release,
|
|
}
|
|
|
|
go func() {
|
|
defer close(s.recvDone)
|
|
for {
|
|
resp, err := stream.Recv()
|
|
if err != nil {
|
|
if !errors.Is(err, io.EOF) && streamCtx.Err() == nil {
|
|
xlog.Warn("live transcription stream ended unexpectedly", "error", err)
|
|
s.recvErr = err
|
|
}
|
|
return
|
|
}
|
|
ev := liveEventFromProto(resp)
|
|
if ev.Delta == "" && !ev.Eou && !ev.Eob && len(ev.Words) == 0 && ev.Final == nil {
|
|
continue // duplicate ready ack / keep-alive: nothing to deliver
|
|
}
|
|
s.trace.observe(ev)
|
|
onEvent(ev)
|
|
}
|
|
}()
|
|
|
|
return s, nil
|
|
}
|
|
|
|
func liveEventFromProto(r *proto.TranscriptLiveResponse) LiveTranscriptionEvent {
|
|
ev := LiveTranscriptionEvent{
|
|
Delta: r.GetDelta(),
|
|
Eou: r.GetEou(),
|
|
Eob: r.GetEob(),
|
|
}
|
|
for _, w := range r.GetWords() {
|
|
ev.Words = append(ev.Words, schema.TranscriptionWord{
|
|
Start: time.Duration(w.Start),
|
|
End: time.Duration(w.End),
|
|
Text: w.Text,
|
|
Speaker: w.Speaker,
|
|
})
|
|
}
|
|
for _, s := range r.GetSpeakers() {
|
|
ev.Speakers = append(ev.Speakers, LiveSpeakerSegment{
|
|
Speaker: s.GetSpeaker(),
|
|
Name: s.GetName(),
|
|
Start: time.Duration(s.GetStart()).Seconds(),
|
|
End: time.Duration(s.GetEnd()).Seconds(),
|
|
})
|
|
}
|
|
for _, s := range r.GetSounds() {
|
|
ev.Sounds = append(ev.Sounds, LiveSoundEvent{
|
|
Label: s.GetLabel(),
|
|
Index: int(s.GetIndex()),
|
|
Peak: s.GetPeak(),
|
|
Start: time.Duration(s.GetStart()).Seconds(),
|
|
End: time.Duration(s.GetEnd()).Seconds(),
|
|
})
|
|
}
|
|
if r.GetFinalResult() != nil {
|
|
ev.Final = transcriptResultFromProto(r.GetFinalResult())
|
|
}
|
|
return ev
|
|
}
|