Sync remote URLs: make GitHub primary, Gitea backup

Dockerfile

@@ -37,7 +37,7 @@ RUN pip install --no-cache-dir torch torchvision torchaudio \
 
 # Install faster-qwen3-tts and server dependencies
 RUN pip install --no-cache-dir -e ".[demo]"
-RUN pip install --no-cache-dir pydub soundfile uvicorn fastapi
+RUN pip install --no-cache-dir pydub soundfile uvicorn fastapi qwen-asr
 
 EXPOSE 8000
 

README.md

@@ -210,7 +210,8 @@ The streaming service on port `8023` uses the same generated `config/voices.json
 | `model` | string | `tts-1` | Kept for OpenAI compatibility |
 | `input` | string | required | Text to synthesize |
 | `voice` | string | first configured voice | Voice ID from the selected service |
-| `response_format` | string | `wav` | `wav`, `pcm`, or `mp3` |
+| `response_format` | string | `wav` | `wav`, `pcm`, `mp3`, or `zip` (for timestamps) |
+| `speed` | float | 1.0 | Scales audio tempo via ffmpeg |
 | `language` | string | voice config | Per-request override for VoiceDesign/CustomVoice |
 | `instruct` | string | voice config | Per-request style override for VoiceDesign/CustomVoice |
 | `max_new_tokens` | int | server default | Per-request generation length override |
@@ -343,6 +344,12 @@ The first request after container startup can be slower because CUDA graph captu
 
 ## Changelog
 
+### v6.7 — 2026-06-26
+**Feature: Native Speed Control and Word-Level Timestamps**
+- **Speed Parameter:** The `speed` parameter in the OpenAI `SpeechRequest` schema is now fully supported. Audio tempo is natively adjusted using `ffmpeg` without affecting pitch, and works for both streaming and non-streaming responses.
+- **Word-Level Timestamps:** Added support for a new `response_format: "zip"`. When requested, the server automatically lazy-loads the `Qwen3-ForcedAligner-0.6B` model to generate word-level timestamps (`timer.json`) and returns it alongside the audio in a compressed zip file.
+- **Input Sanitization:** Automatically strips leading and trailing whitespace from input text to fix a bug where excessive blank space caused the tokenizer to stutter and repeat words.
+
 ### v6.6 — 2026-06-21
 **Feature: Eager Background Precomputation of Speaker Embeddings**
 - The server now automatically precomputes all missing `.pt` files in the background immediately after startup.

config/run_customvoice_server.py

@@ -37,7 +37,24 @@ SAMPLE_RATE = 24000
 DEFAULT_MAX_NEW_TOKENS = 2048
 _model_lock = threading.Lock()
 _load_model_kwargs = None
+aligner_model = None
 
+def _get_aligner():
+    global aligner_model
+    if aligner_model is None:
+        try:
+            from qwen_asr import Qwen3ForcedAligner
+            import torch
+        except ImportError:
+            raise HTTPException(status_code=500, detail="qwen-asr is not installed. Run: pip install qwen-asr")
+        logger.info("Loading Qwen3-ForcedAligner-0.6B...")
+        aligner_model = Qwen3ForcedAligner.from_pretrained(
+            "Qwen/Qwen3-ForcedAligner-0.6B", 
+            dtype=torch.bfloat16, 
+            device_map="cuda"
+        )
+        logger.info("Aligner loaded.")
+    return aligner_model
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
@@ -87,7 +104,7 @@ class SpeechRequest(BaseModel):
     model: str = "tts-1"
     input: str
     voice: str = "Ryan"
-    response_format: str = "wav"
+    response_format: str = "wav"  # wav | pcm | mp3 | zip
     speed: float = 1.0
     language: Optional[str] = None
     instruct: Optional[str] = None
@@ -141,20 +158,58 @@ def _request_generation_params(req: SpeechRequest, voice_cfg: dict) -> dict:
     }
 
 
-async def _stream_chunks(params: dict):
+async def _stream_chunks(params: dict, speed: float):
     q: queue.Queue = queue.Queue()
     done = object()
 
     def producer():
+        process = None
+        if speed != 1.0:
+            import subprocess
+            cmd = [
+                "ffmpeg", "-y", "-loglevel", "error",
+                "-f", "s16le", "-ar", str(SAMPLE_RATE), "-ac", "1", "-i", "pipe:0",
+                "-filter:a", f"atempo={speed}",
+                "-f", "s16le", "-ar", str(SAMPLE_RATE), "-ac", "1", "pipe:1"
+            ]
+            process = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
+            
+            def ffmpeg_reader():
                 try:
-            with _model_lock:
-                for chunk, _sr, _timing in tts_model.generate_custom_voice_streaming(**params):
-                    q.put(chunk)
-        except Exception as exc:
-            q.put(exc)
+                    while True:
+                        out = process.stdout.read(4096)
+                        if not out:
+                            break
+                        q.put(out)
+                except Exception as e:
+                    q.put(e)
                 finally:
                     q.put(done)
                     
+            import threading
+            threading.Thread(target=ffmpeg_reader, daemon=True).start()
+
+        try:
+            with _model_lock:
+                for chunk, _sr, _timing in tts_model.generate_custom_voice_streaming(**params):
+                    raw = _to_pcm16(chunk)
+                    if process:
+                        process.stdin.write(raw)
+                        process.stdin.flush()
+                    else:
+                        q.put(raw)
+        except Exception as exc:
+            q.put(exc)
+        finally:
+            if process:
+                try:
+                    process.stdin.close()
+                except Exception:
+                    pass
+            else:
+                q.put(done)
+
+    import threading
     threading.Thread(target=producer, daemon=True).start()
     loop = asyncio.get_event_loop()
     while True:
@@ -163,7 +218,7 @@ async def _stream_chunks(params: dict):
             break
         if isinstance(item, Exception):
             raise item
-        yield _to_pcm16(item)
+        yield item
 
 
 @app.get("/health")
@@ -175,18 +230,19 @@ async def health():
 async def create_speech(req: SpeechRequest):
     if tts_model is None:
         raise HTTPException(status_code=503, detail="Model not loaded")
-    if not req.input.strip():
+    req.input = req.input.strip()
+    if not req.input:
         raise HTTPException(status_code=400, detail="'input' text is empty")
 
     voice_cfg = resolve_voice(req.voice)
     params = _request_generation_params(req, voice_cfg)
     fmt = req.response_format.lower()
 
-    content_types = {"wav": "audio/wav", "pcm": "audio/pcm", "mp3": "audio/mpeg"}
+    content_types = {"wav": "audio/wav", "pcm": "audio/pcm", "mp3": "audio/mpeg", "zip": "application/zip"}
     if fmt not in content_types:
         raise HTTPException(status_code=400, detail=f"Unsupported format: {fmt!r}")
 
-    if fmt == "mp3":
+    if fmt in ("mp3", "zip"):
         loop = asyncio.get_event_loop()
 
         def generate():
@@ -195,12 +251,46 @@ async def create_speech(req: SpeechRequest):
 
         audio_arrays, sr = await loop.run_in_executor(None, generate)
         audio = audio_arrays[0] if audio_arrays else np.zeros(1, dtype=np.float32)
+        
+        if req.speed != 1.0:
+            import subprocess
+            cmd = [
+                "ffmpeg", "-y", "-loglevel", "error",
+                "-f", "f32le", "-ar", str(sr), "-ac", "1", "-i", "pipe:0",
+                "-filter:a", f"atempo={req.speed}",
+                "-f", "f32le", "-ar", str(sr), "-ac", "1", "pipe:1"
+            ]
+            process = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
+            process.stdin.write(audio.tobytes())
+            process.stdin.close()
+            out = process.stdout.read()
+            audio = np.frombuffer(out, dtype=np.float32)
+
+        if fmt == "zip":
+            def _align():
+                aligner = _get_aligner()
+                res = aligner.align(audio=(audio, sr), text=req.input, language=voice_cfg.get("language", "Auto"))
+                import dataclasses
+                return [dataclasses.asdict(x) for x in res]
+            
+            align_data = await loop.run_in_executor(None, _align)
+            
+            import zipfile
+            import io
+            mp3_bytes = _to_mp3_bytes(audio, sr)
+            zip_buf = io.BytesIO()
+            with zipfile.ZipFile(zip_buf, "w", zipfile.ZIP_DEFLATED) as zf:
+                zf.writestr("audio.mp3", mp3_bytes)
+                zf.writestr("timer.json", json.dumps(align_data, ensure_ascii=False))
+                
+            return Response(content=zip_buf.getvalue(), media_type=content_types[fmt])
+
         return Response(content=_to_mp3_bytes(audio, sr), media_type="audio/mpeg")
 
     async def audio_stream():
         if fmt == "wav":
             yield _wav_header(SAMPLE_RATE)
-        async for raw in _stream_chunks(params):
+        async for raw in _stream_chunks(params, req.speed):
             yield raw
 
     return StreamingResponse(audio_stream(), media_type=content_types[fmt])

config/run_voicedesign_server.py

@@ -36,7 +36,24 @@ SAMPLE_RATE = 24000
 DEFAULT_MAX_NEW_TOKENS = 2048
 _model_lock = threading.Lock()
 _load_model_kwargs = None
+aligner_model = None
 
+def _get_aligner():
+    global aligner_model
+    if aligner_model is None:
+        try:
+            from qwen_asr import Qwen3ForcedAligner
+            import torch
+        except ImportError:
+            raise HTTPException(status_code=500, detail="qwen-asr is not installed. Run: pip install qwen-asr")
+        logger.info("Loading Qwen3-ForcedAligner-0.6B...")
+        aligner_model = Qwen3ForcedAligner.from_pretrained(
+            "Qwen/Qwen3-ForcedAligner-0.6B", 
+            dtype=torch.bfloat16, 
+            device_map="cuda"
+        )
+        logger.info("Aligner loaded.")
+    return aligner_model
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
@@ -90,7 +107,7 @@ class SpeechRequest(BaseModel):
     model: str = "tts-1"
     input: str
     voice: str = "vd_british_male"
-    response_format: str = "wav"
+    response_format: str = "wav"  # wav | pcm | mp3 | zip
     speed: float = 1.0
     language: Optional[str] = None
     instruct: Optional[str] = None
@@ -151,20 +168,58 @@ def _request_generation_params(req: SpeechRequest, voice_cfg: dict) -> dict:
     }
 
 
-async def _stream_chunks(params: dict):
+async def _stream_chunks(params: dict, speed: float):
     q: queue.Queue = queue.Queue()
     _DONE = object()
 
     def producer():
+        process = None
+        if speed != 1.0:
+            import subprocess
+            cmd = [
+                "ffmpeg", "-y", "-loglevel", "error",
+                "-f", "s16le", "-ar", str(SAMPLE_RATE), "-ac", "1", "-i", "pipe:0",
+                "-filter:a", f"atempo={speed}",
+                "-f", "s16le", "-ar", str(SAMPLE_RATE), "-ac", "1", "pipe:1"
+            ]
+            process = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
+            
+            def ffmpeg_reader():
                 try:
-            with _model_lock:
-                for chunk, _sr, _timing in tts_model.generate_voice_design_streaming(**params):
-                    q.put(chunk)
-        except Exception as exc:
-            q.put(exc)
+                    while True:
+                        out = process.stdout.read(4096)
+                        if not out:
+                            break
+                        q.put(out)
+                except Exception as e:
+                    q.put(e)
                 finally:
                     q.put(_DONE)
                     
+            import threading
+            threading.Thread(target=ffmpeg_reader, daemon=True).start()
+
+        try:
+            with _model_lock:
+                for chunk, _sr, _timing in tts_model.generate_voice_design_streaming(**params):
+                    raw = _to_pcm16(chunk)
+                    if process:
+                        process.stdin.write(raw)
+                        process.stdin.flush()
+                    else:
+                        q.put(raw)
+        except Exception as exc:
+            q.put(exc)
+        finally:
+            if process:
+                try:
+                    process.stdin.close()
+                except Exception:
+                    pass
+            else:
+                q.put(_DONE)
+
+    import threading
     threading.Thread(target=producer, daemon=True).start()
     loop = asyncio.get_event_loop()
     while True:
@@ -173,7 +228,7 @@ async def _stream_chunks(params: dict):
             break
         if isinstance(item, Exception):
             raise item
-        yield _to_pcm16(item)
+        yield item
 
 
 # ---------------------------------------------------------------------------
@@ -189,30 +244,65 @@ async def health():
 async def create_speech(req: SpeechRequest):
     if tts_model is None:
         raise HTTPException(status_code=503, detail="Model not loaded")
-    if not req.input.strip():
+    req.input = req.input.strip()
+    if not req.input:
         raise HTTPException(status_code=400, detail="'input' text is empty")
 
     voice_cfg = resolve_voice(req.voice)
     params = _request_generation_params(req, voice_cfg)
     fmt = req.response_format.lower()
 
-    _CONTENT_TYPES = {"wav": "audio/wav", "pcm": "audio/pcm", "mp3": "audio/mpeg"}
+    _CONTENT_TYPES = {"wav": "audio/wav", "pcm": "audio/pcm", "mp3": "audio/mpeg", "zip": "application/zip"}
     if fmt not in _CONTENT_TYPES:
         raise HTTPException(status_code=400, detail=f"Unsupported format: {fmt!r}")
 
-    if fmt == "mp3":
+    if fmt in ("mp3", "zip"):
         loop = asyncio.get_event_loop()
         def _gen():
             with _model_lock:
                 return tts_model.generate_voice_design(**params)
         audio_arrays, sr = await loop.run_in_executor(None, _gen)
         audio = audio_arrays[0] if audio_arrays else np.zeros(1, dtype=np.float32)
+        
+        if req.speed != 1.0:
+            import subprocess
+            cmd = [
+                "ffmpeg", "-y", "-loglevel", "error",
+                "-f", "f32le", "-ar", str(sr), "-ac", "1", "-i", "pipe:0",
+                "-filter:a", f"atempo={req.speed}",
+                "-f", "f32le", "-ar", str(sr), "-ac", "1", "pipe:1"
+            ]
+            process = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
+            process.stdin.write(audio.tobytes())
+            process.stdin.close()
+            out = process.stdout.read()
+            audio = np.frombuffer(out, dtype=np.float32)
+
+        if fmt == "zip":
+            def _align():
+                aligner = _get_aligner()
+                res = aligner.align(audio=(audio, sr), text=req.input, language=voice_cfg.get("language", "Auto"))
+                import dataclasses
+                return [dataclasses.asdict(x) for x in res]
+            
+            align_data = await loop.run_in_executor(None, _align)
+            
+            import zipfile
+            import io
+            mp3_bytes = _to_mp3_bytes(audio, sr)
+            zip_buf = io.BytesIO()
+            with zipfile.ZipFile(zip_buf, "w", zipfile.ZIP_DEFLATED) as zf:
+                zf.writestr("audio.mp3", mp3_bytes)
+                zf.writestr("timer.json", json.dumps(align_data, ensure_ascii=False))
+                
+            return Response(content=zip_buf.getvalue(), media_type=_CONTENT_TYPES[fmt])
+
         return Response(content=_to_mp3_bytes(audio, sr), media_type="audio/mpeg")
 
     async def audio_stream():
         if fmt == "wav":
             yield _wav_header(SAMPLE_RATE)
-        async for raw in _stream_chunks(params):
+        async for raw in _stream_chunks(params, req.speed):
             yield raw
 
     return StreamingResponse(audio_stream(), media_type=_CONTENT_TYPES[fmt])

patches/openai_server.patch

@@ -1,6 +1,8 @@
---- /tmp/upstream_openai_server.py	2026-06-21 14:29:34.858114787 +0200
-+++ build/examples/openai_server.py	2026-06-21 14:26:33.557536313 +0200
-@@ -36,6 +36,7 @@
+diff --git a/examples/openai_server.py b/examples/openai_server.py
+index 2199e14..d10065f 100644
+--- a/examples/openai_server.py
++++ b/examples/openai_server.py
+@@ -36,6 +36,7 @@ API usage:
  """
  import argparse
  import asyncio
@@ -8,7 +10,7 @@
  import io
  import json
  import logging
-@@ -66,10 +67,29 @@
+@@ -66,9 +67,46 @@ app = FastAPI(title="faster-qwen3-tts OpenAI-compatible API")
  
  tts_model = None
  voices: dict = {}
@@ -17,7 +19,25 @@
  default_voice: Optional[str] = None
  SAMPLE_RATE = 24000  # updated once the model loads
  _model_lock = threading.Lock()  # prevent concurrent GPU inference
- 
++aligner_model = None
++
++def _get_aligner():
++    global aligner_model
++    if aligner_model is None:
++        try:
++            from qwen_asr import Qwen3ForcedAligner
++            import torch
++        except ImportError:
++            raise HTTPException(status_code=500, detail="qwen-asr is not installed. Run: pip install qwen-asr")
++        logger.info("Loading Qwen3-ForcedAligner-0.6B...")
++        aligner_model = Qwen3ForcedAligner.from_pretrained(
++            "Qwen/Qwen3-ForcedAligner-0.6B", 
++            dtype=torch.bfloat16, 
++            device_map="cuda"
++        )
++        logger.info("Aligner loaded.")
++    return aligner_model
++
 +
 +def _voice_seed(voice_name: str) -> int:
 +    """Return a stable per-voice seed derived from the voice name.
@@ -34,11 +54,21 @@
 +    if torch.cuda.is_available():
 +        torch.cuda.manual_seed_all(seed)
 +
-+
+ 
  # ---------------------------------------------------------------------------
  # Request / response models
+@@ -79,8 +117,8 @@ class SpeechRequest(BaseModel):
+     model: str = "tts-1"
+     input: str
+     voice: str = "alloy"
+-    response_format: str = "wav"  # wav | pcm | mp3
+-    speed: float = 1.0           # accepted but not yet applied
++    response_format: str = "wav"  # wav | pcm | mp3 | zip
++    speed: float = 1.0           # scales audio tempo
+ 
+ 
  # ---------------------------------------------------------------------------
-@@ -145,6 +165,21 @@
+@@ -145,6 +183,21 @@ def _to_mp3_bytes(pcm: np.ndarray, sample_rate: int) -> bytes:
  
  def resolve_voice(voice_name: str) -> dict:
      """Return voice config dict or fall back to default, else raise 400."""
@@ -60,7 +90,7 @@
      if voice_name in voices:
          return voices[voice_name]
      if default_voice and default_voice in voices:
-@@ -168,7 +203,47 @@
+@@ -168,7 +221,47 @@ def resolve_voice(voice_name: str) -> dict:
  # ---------------------------------------------------------------------------
  
  
@@ -105,12 +135,39 @@
 +        return None
 +
 +
-+async def _stream_chunks(voice_cfg: dict, text: str, voice_name: str) -> AsyncGenerator[bytes, None]:
++async def _stream_chunks(voice_cfg: dict, text: str, voice_name: str, speed: float) -> AsyncGenerator[bytes, None]:
      """
      Run generate_voice_clone_streaming in a background thread and yield
      raw PCM bytes for each chunk as they arrive.
-@@ -179,13 +254,19 @@
+@@ -177,21 +270,63 @@ async def _stream_chunks(voice_cfg: dict, text: str) -> AsyncGenerator[bytes, No
+     _DONE = object()
+ 
      def producer():
++        process = None
++        if speed != 1.0:
++            import subprocess
++            cmd = [
++                "ffmpeg", "-y", "-loglevel", "error",
++                "-f", "s16le", "-ar", str(SAMPLE_RATE), "-ac", "1", "-i", "pipe:0",
++                "-filter:a", f"atempo={speed}",
++                "-f", "s16le", "-ar", str(SAMPLE_RATE), "-ac", "1", "pipe:1"
++            ]
++            process = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
++            
++            def ffmpeg_reader():
++                try:
++                    while True:
++                        out = process.stdout.read(4096)
++                        if not out:
++                            break
++                        q.put(out)
++                except Exception as e:
++                    q.put(e)
++                finally:
++                    q.put(_DONE)
++                    
++            threading.Thread(target=ffmpeg_reader, daemon=True).start()
++
          try:
              with _model_lock:
 +                _seed_rng(voice_cfg.get("seed", _voice_seed(voice_name)))
@@ -129,9 +186,65 @@
 +                    top_k=voice_cfg.get("top_k", 50),
 +                    top_p=voice_cfg.get("top_p", 0.9),
                  ):
-                     q.put(chunk)
+-                    q.put(chunk)
++                    raw = _to_pcm16(chunk)
++                    if process:
++                        process.stdin.write(raw)
++                        process.stdin.flush()
++                    else:
++                        q.put(raw)
          except Exception as exc:
-@@ -244,11 +325,18 @@
+             q.put(exc)
+         finally:
+-            q.put(_DONE)
++            if process:
++                try:
++                    process.stdin.close()
++                except Exception:
++                    pass
++            else:
++                q.put(_DONE)
+ 
+     thread = threading.Thread(target=producer, daemon=True)
+     thread.start()
+@@ -203,7 +338,7 @@ async def _stream_chunks(voice_cfg: dict, text: str) -> AsyncGenerator[bytes, No
+             break
+         if isinstance(item, Exception):
+             raise item
+-        yield _to_pcm16(item)
++        yield item
+ 
+ 
+ # ---------------------------------------------------------------------------
+@@ -220,7 +355,8 @@ async def health():
+ async def create_speech(req: SpeechRequest):
+     if tts_model is None:
+         raise HTTPException(status_code=503, detail="Model not loaded")
+-    if not req.input.strip():
++    req.input = req.input.strip()
++    if not req.input:
+         raise HTTPException(status_code=400, detail="'input' text is empty")
+ 
+     voice_cfg = resolve_voice(req.voice)
+@@ -230,36 +366,77 @@ async def create_speech(req: SpeechRequest):
+         "wav": "audio/wav",
+         "pcm": "audio/pcm",
+         "mp3": "audio/mpeg",
++        "zip": "application/zip",
+     }
+     if fmt not in _CONTENT_TYPES:
+         raise HTTPException(
+             status_code=400,
+-            detail=f"response_format {fmt!r} not supported. Use: wav, pcm, mp3",
++            detail=f"response_format {fmt!r} not supported. Use: wav, pcm, mp3, zip",
+         )
+     content_type = _CONTENT_TYPES[fmt]
+ 
+-    # --- MP3: generate all audio, then encode (non-streaming) ---
+-    if fmt == "mp3":
++    # --- MP3 / ZIP: generate all audio, then encode (non-streaming) ---
++    if fmt in ("mp3", "zip"):
+         loop = asyncio.get_event_loop()
  
          def _generate():
              with _model_lock:
@@ -151,16 +264,52 @@
                  )
  
          audio_arrays, sr = await loop.run_in_executor(None, _generate)
-@@ -259,7 +347,7 @@
+         audio = audio_arrays[0] if audio_arrays else np.zeros(1, dtype=np.float32)
++        
++        if req.speed != 1.0:
++            import subprocess
++            cmd = [
++                "ffmpeg", "-y", "-loglevel", "error",
++                "-f", "f32le", "-ar", str(sr), "-ac", "1", "-i", "pipe:0",
++                "-filter:a", f"atempo={req.speed}",
++                "-f", "f32le", "-ar", str(sr), "-ac", "1", "pipe:1"
++            ]
++            process = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE)
++            process.stdin.write(audio.tobytes())
++            process.stdin.close()
++            out = process.stdout.read()
++            audio = np.frombuffer(out, dtype=np.float32)
++
++        if fmt == "zip":
++            def _align():
++                aligner = _get_aligner()
++                res = aligner.align(audio=(audio, sr), text=req.input, language=voice_cfg.get("language", "Auto"))
++                import dataclasses
++                return [dataclasses.asdict(x) for x in res]
++            
++            align_data = await loop.run_in_executor(None, _align)
++            
++            import zipfile
++            mp3_bytes = _to_mp3_bytes(audio, sr)
++            zip_buf = io.BytesIO()
++            with zipfile.ZipFile(zip_buf, "w", zipfile.ZIP_DEFLATED) as zf:
++                zf.writestr("audio.mp3", mp3_bytes)
++                zf.writestr("timer.json", json.dumps(align_data, ensure_ascii=False))
++                
++            return Response(content=zip_buf.getvalue(), media_type=content_type)
++
+         return Response(content=_to_mp3_bytes(audio, sr), media_type=content_type)
+ 
+     # --- WAV / PCM: stream chunks as they are generated ---
      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):
++        async for raw_chunk in _stream_chunks(voice_cfg, req.input, req.voice, req.speed):
              yield raw_chunk
  
      return StreamingResponse(audio_stream(), media_type=content_type)
-@@ -306,16 +394,20 @@
+@@ -306,16 +483,20 @@ def _parse_args():
      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)")
@@ -182,7 +331,7 @@
          with open(args.voices) as f:
              voices = json.load(f)
          default_voice = next(iter(voices))
-@@ -344,6 +436,7 @@
+@@ -344,6 +525,7 @@ def main():
          args.model,
          device=args.device,
          dtype=torch.bfloat16,