feat: fully automate speaker embedding generation (v6.4)

README.md

@@ -138,6 +138,7 @@ python config/auto_transcribe.py --api-url http://localhost:8010/v1/audio/transc
 
 `config/generate_voices.py` runs on container startup and creates `config/voices.json` from your speaker files.
 
+When you start using a new voice for the first time, the server will automatically do the heavy lifting to extract the voice's acoustic fingerprint (a "speaker embedding") and save it as a `.pt` file in the `config/speakers/` directory. Future requests will instantly load this `.pt` file instead of re-analyzing the audio, which dramatically speeds up Time To First Audio (TTFA).
 ## VoiceDesign voices
 
 VoiceDesign does not need reference audio. Define reusable voice personalities in `config/voicedesign_voices.json`:
@@ -340,6 +341,14 @@ The first request after container startup can be slower because CUDA graph captu
 
 ## Changelog
 
+### v6.4 — 2026-06-20
+**Feature: Fully Automated Speaker Embeddings (.pt files)**
+
+- Integrated speaker embedding extraction directly into the API server (`openai_server.py`). 
+- When a new voice is requested for the first time, the server will automatically compute the speaker embedding and save it as a `.pt` file in `/config/speakers/`.
+- Future requests for the same voice automatically use the `.pt` file instead of recalculating the prompt from the `.wav` or `.mp3` reference audio.
+- This provides the massive TTFA speedup of precomputed embeddings without requiring any manual scripting or configuration.
+
 ### v6.3 — 2026-06-20
 **Feature: Precomputed Speaker Embeddings (.pt files)**
 

patches/openai_server.patch

@@ -1,25 +1,63 @@
-diff --git a/examples/openai_server.py b/examples/openai_server.py
+--- /tmp/upstream_openai_server.py	2026-06-20 20:43:53.380341702 +0200
-index 61047ea..2f95bbd 100644
++++ build/examples/openai_server.py	2026-06-20 20:52:17.725652965 +0200
---- a/examples/openai_server.py
+@@ -145,6 +145,7 @@
-+++ b/examples/openai_server.py
+ 
-@@ -169,6 +169,16 @@ def resolve_voice(voice_name: str) -> dict:
+ def resolve_voice(voice_name: str) -> dict:
+     """Return voice config dict or fall back to default, else raise 400."""
++    voice_name = voice_name.strip()
+     if voice_name in voices:
+         return voices[voice_name]
+     if default_voice and default_voice in voices:
+@@ -168,7 +169,47 @@
  # ---------------------------------------------------------------------------
  
  
-+def _load_voice_clone_prompt(voice_cfg: dict):
+-async def _stream_chunks(voice_cfg: dict, text: str) -> AsyncGenerator[bytes, None]:
++def _load_voice_clone_prompt(voice_cfg: dict, voice_name: str, tts_model):
 +    spk_emb_path = voice_cfg.get("speaker_embeddings") or voice_cfg.get("speaker embeddings")
 +    if spk_emb_path and os.path.isfile(spk_emb_path):
 +        try:
 +            return torch.load(spk_emb_path, map_location="cpu", weights_only=False)
 +        except Exception as e:
 +            logger.error("Failed to load speaker embeddings from %s: %s", spk_emb_path, e)
-+    return None
++            
++    # Auto-generate if missing
++    ref_audio = voice_cfg.get("ref_audio")
++    if not ref_audio or not os.path.isfile(ref_audio):
++        return None
++        
++    logger.info("Precomputing and saving speaker embedding for voice %r...", voice_name)
++    try:
++        ref_text = voice_cfg.get("ref_text", "")
++        # generate prompt using the model's built-in helper
++        prompt_items = tts_model.model.create_voice_clone_prompt(ref_audio, [ref_text])
++        vcp = tts_model.model._prompt_items_to_voice_clone_prompt(prompt_items)
++        
++        # save it to the speakers directory
++        # If the file path is already in voice_cfg but doesn't exist, use that, otherwise generate a path
++        if spk_emb_path and not os.path.exists(spk_emb_path) and spk_emb_path.endswith('.pt'):
++            pt_path = spk_emb_path
++        else:
++            pt_path = f"/config/speakers/{voice_name}.pt"
++            
++        os.makedirs(os.path.dirname(pt_path), exist_ok=True)
++        torch.save(vcp, pt_path)
++        logger.info("Saved speaker embedding to %s", pt_path)
++        
++        # update in memory so future requests skip generating
++        voice_cfg["speaker_embeddings"] = pt_path
++        
++        return vcp
++    except Exception as e:
++        logger.error("Failed to precompute speaker embedding: %s", e)
++        return None
 +
 +
- async def _stream_chunks(voice_cfg: dict, text: str) -> AsyncGenerator[bytes, None]:
++async def _stream_chunks(voice_cfg: dict, text: str, voice_name: str) -> AsyncGenerator[bytes, None]:
      """
      Run generate_voice_clone_streaming in a background thread and yield
-@@ -183,11 +193,15 @@ async def _stream_chunks(voice_cfg: dict, text: str) -> AsyncGenerator[bytes, No
+     raw PCM bytes for each chunk as they arrive.
+@@ -182,10 +223,15 @@
                  for chunk, _sr, _timing in tts_model.generate_voice_clone_streaming(
                      text=text,
                      language=voice_cfg.get("language", "Auto"),
@@ -27,9 +65,9 @@ index 61047ea..2f95bbd 100644
 +                    ref_audio=voice_cfg.get("ref_audio"),
                      ref_text=voice_cfg.get("ref_text", ""),
                      chunk_size=voice_cfg.get("chunk_size", 12),
-                     instruct=voice_cfg.get("instruct"),
 -                    non_streaming_mode=False,
-+                    voice_clone_prompt=_load_voice_clone_prompt(voice_cfg),
++                    instruct=voice_cfg.get("instruct"),
++                    voice_clone_prompt=_load_voice_clone_prompt(voice_cfg, voice_name, tts_model),
 +                    non_streaming_mode=True,
 +                    temperature=voice_cfg.get("temperature", 0.8),
 +                    top_k=voice_cfg.get("top_k", 50),
@@ -37,15 +75,15 @@ index 61047ea..2f95bbd 100644
                  ):
                      q.put(chunk)
          except Exception as exc:
-@@ -249,9 +263,14 @@ async def create_speech(req: SpeechRequest):
+@@ -247,8 +293,14 @@
                  return tts_model.generate_voice_clone(
                      text=req.input,
                      language=voice_cfg.get("language", "Auto"),
 -                    ref_audio=voice_cfg["ref_audio"],
 +                    ref_audio=voice_cfg.get("ref_audio"),
                      ref_text=voice_cfg.get("ref_text", ""),
-                     instruct=voice_cfg.get("instruct"),
++                    instruct=voice_cfg.get("instruct"),
-+                    voice_clone_prompt=_load_voice_clone_prompt(voice_cfg),
++                    voice_clone_prompt=_load_voice_clone_prompt(voice_cfg, req.voice, tts_model),
 +                    non_streaming_mode=True,
 +                    temperature=voice_cfg.get("temperature", 0.8),
 +                    top_k=voice_cfg.get("top_k", 50),
@@ -53,3 +91,28 @@ index 61047ea..2f95bbd 100644
                  )
  
          audio_arrays, sr = await loop.run_in_executor(None, _generate)
+@@ -259,7 +311,7 @@
+     async def audio_stream():
+         if fmt == "wav":
+             yield _wav_header(SAMPLE_RATE)  # stream with unknown data length
+-        async for raw_chunk in _stream_chunks(voice_cfg, req.input):
++        async for raw_chunk in _stream_chunks(voice_cfg, req.input, req.voice):
+             yield raw_chunk
+ 
+     return StreamingResponse(audio_stream(), media_type=content_type)
+@@ -306,6 +358,7 @@
+     p.add_argument("--host", default="0.0.0.0", help="Bind host (default: 0.0.0.0)")
+     p.add_argument("--port", type=int, default=8000, help="Bind port (default: 8000)")
+     p.add_argument("--device", default="cuda", help="Torch device (default: cuda)")
++    p.add_argument("--max-seq-len", type=int, default=4096, help="Max sequence length for CUDA graph static cache (default: 4096)")
+     return p.parse_args()
+ 
+ 
+@@ -344,6 +397,7 @@
+         args.model,
+         device=args.device,
+         dtype=torch.bfloat16,
++        max_seq_len=args.max_seq_len,
+     )
+     SAMPLE_RATE = tts_model.sample_rate
+     logger.info("Model ready. Sample rate: %d Hz", SAMPLE_RATE)