first

## Motivation

Human evaluations consistently prefer the English model at temperature
**0.3** over the current default **0.7**, and the change is free on
every objective axis we measured.

.dbg/trae-debug-log-swagger-audio-play.ndjson

@@ -0,0 +1,4 @@
+{"sessionId": "swagger-audio-play", "runId": "pre-fix", "hypothesisId": "A", "location": "server.py:/tts", "msg": "[DEBUG] TTS request received", "data": {"format": "mp3", "voice_url": "lola", "text_length": 1500}, "ts": 1784807181342}
+{"sessionId": "swagger-audio-play", "runId": "pre-fix", "hypothesisId": "B", "location": "server.py:/tts", "msg": "[DEBUG] Audio payload prepared", "data": {"media_type": "audio/mpeg", "sample_rate": 24000, "audio_bytes_len": 767832, "first_16_bytes_hex": "fff384c400000000000000000058696e"}, "ts": 1784807884454}
+{"sessionId": "swagger-audio-play", "runId": "pre-fix", "hypothesisId": "C", "location": "server.py:/tts", "msg": "[DEBUG] Response branch selected", "data": {"format": "mp3", "media_type": "audio/mpeg"}, "ts": 1784807884455}
+{"sessionId": "swagger-audio-play", "runId": "pre-fix", "hypothesisId": "A", "location": "server.py:/tts", "msg": "[DEBUG] TTS request received", "data": {"format": "mp3", "voice_url": "lola", "text_length": 1290}, "ts": 1784807884459}

README.md

@@ -162,7 +162,6 @@ You can check out the [Python API documentation](https://kyutai-labs.github.io/p
 At the moment, we do not support (but would love pull requests adding):
 
 - [Adding silence in the text input to generate pauses.](https://github.com/kyutai-labs/pocket-tts/issues/6)
-- [Quantization to run the computation in int8.](https://github.com/kyutai-labs/pocket-tts/issues/7)
 
 We tried running this TTS model on the GPU but did not observe a speedup compared to CPU execution,
 notably because we use a batch size of 1 and a very small model.
@@ -196,7 +195,8 @@ We don't have official support for this yet, but you can try out one of these co
 - [pocket-reader](https://github.com/lukasmwerner/pocket-reader) by @lukasmwerner- Browser screen reader
 - [pocket-tts-wyoming](https://github.com/ikidd/pocket-tts-wyoming) by @ikidd - Docker container for pocket-tts using Wyoming protocol, ready for Home Assistant Voice use.
 - [Sonorus](https://www.nexusmods.com/hogwartslegacy/mods/2409) by @KevinAHM - Talk to any named character in Hogwarts Legacy with their original voice.
-- [Mac pocket-tts](https://github.com/slaughters85j/pocket-tts) by @slaughters85j - Mac Desktop App + macOS Quick Action
+- [Native macOS App](https://github.com/slaughters85j/pocket-tts-macos) by @slaughters85j - Native macOS app, Python-free. Runs Pocket-TTS via Core ML, fully on-device. Includes signed and notarized .app releases.
+- [Electron macOS App](https://github.com/slaughters85j/pocket-tts) by @slaughters85j - Electron Mac Desktop App + macOS Quick Action
 - [pocket-tts-openai_streaming_server](https://github.com/teddybear082/pocket-tts-openai_streaming_server) by @teddybear082 - OpenAI-compatible streaming server, dockerized and with an `.exe` release
 - [pocket-tts-unity](https://github.com/lookbe/pocket-tts-unity) by @lookbe - A Unity 6 integration for Pocket-TTS.
 - [ComfyUI-Pocket-TTS](https://github.com/ai-joe-git/ComfyUI-Pocket-TTS) by @ai-joe-git Lightweight CPU-based Text-to-Speech for ComfyUI
@@ -208,6 +208,8 @@ We don't have official support for this yet, but you can try out one of these co
 - [openclaw-pockettts](https://github.com/dodgyrabbit/openclaw-pockettts) by @dodgyrabbit - A Docker container with the Python implementation but exposed as an OpenAI TTS API for easy integration with OpenClaw.
 - [openclaw-pocketts.cpp](https://github.com/dodgyrabbit/openclaw-pockettts.cpp) by @dodgyrabbit - A Docker container with the PocketTTS.cpp version, packaged for easy integration with OpenClaw.
 - [tts-audiobook-tool](https://github.com/zeropointnine/tts-audiobook-tool) by @zeropointnine - Multi-model audiobook generator with automatic error detection, 48khz upscaling, synced browser reader, stand-alone server-mode.
+- [seshat-tts](https://github.com/scriptriva/seshat-tts) by @scriptriva - Accessibility tool that provides real-time audio synthesis for games and apps. It also features a voice manager capable of cloning voices based on user presets.
+- [LocalVocal.ai](https://localvocal.ai) by @joshwhiton - Fully local conversational voice-harness for Macs with Apple Silicon. Includes voice-activity & turn detection, dictation, voice cloning, CLI to talk to Claude, Codex... and more.
 
 
 ## Prohibited use

docker-bake.hcl

@@ -9,7 +9,7 @@ group "default" {
 target "pocket-tts-server" {
   context = "."
   dockerfile = "Dockerfile"
-  platforms = ["linux/arm64", "linux/amd64"]
+  platforms = ["linux/amd64"]
   tags = [
     "rg.fr-par.scw.cloud/namespace-unruffled-tereshkova/pocket-tts-server:${COMMIT_SHA}"
   ]

docs/API Reference/python-api.md

@@ -37,14 +37,14 @@ The main class for text-to-speech generation.
 
 #### Class Methods
 
-##### `load_model(language=None, config=None, temp=0.7, lsd_decode_steps=1, noise_clamp=None, eos_threshold=-4.0, quantize=False)`
+##### `load_model(language=None, config=None, temp=None, lsd_decode_steps=1, noise_clamp=None, eos_threshold=-4.0, quantize=False)`
 
 Load and return a TTSModel instance with pre-trained weights.
 
 **Parameters:**
 - `language` (str | None): Name of built-in language config to load. Supported values: `"english_2026-01"`, `"english_2026-04"`, `"english"`, `"french_24l"`, `"german_24l"`, `"portuguese_24l"`, `"italian_24l"`, `"spanish_24l"`. If both `language` and `config` are omitted, defaults to `"english"`, which is the same model as `"english_2026-04"`. The "24l" variants are larger models that are not distilled yet and are here only as a preview.
 - `config` (str | None): Path to model config YAML file. Incompatible with `language`.
-- `temp` (float): Sampling temperature for generation (default: 0.7)
+- `temp` (float | None): Sampling temperature for generation. `None` uses the model's recommended default from its config file (`default_temperature`; 0.3 for the English model, 0.7 otherwise).
 - `lsd_decode_steps` (int): Number of generation steps (default: 1)
 - `noise_clamp` (float | None): Maximum value for noise sampling (default: None)
 - `eos_threshold` (float): Threshold for end-of-sequence detection (default: -4.0)

docs/CLI Commands/generate.md

@@ -17,15 +17,15 @@ This will generate a WAV file `./tts_output.wav` with the default text and voice
 ### Core Options
 
 - `--text TEXT`: Text to generate (default: "Hello world! I am Kyutai Pocket TTS. I'm fast enough to run on small CPUs. I hope you'll like me.")
-- `--voice VOICE`: Path to audio conditioning file (voice to clone). Defaults to a built-in voice chosen from the language: `giovanni` (it) for italian, `lola` (es) for spanish, `alba` (en) otherwise. Urls and local paths are supported.
+- `--voice VOICE`: Path to audio conditioning file (voice to clone). Defaults to a built-in voice chosen from the language: `giovanni` (it) for italian, `lola` (es) for spanish, `juergen` (de) for german, `rafael` (pt) for portuguese, `estelle` (fr) for french, `alba` (en) otherwise. Urls and local paths are supported.
 - `--output-path OUTPUT_PATH`: Output path for generated audio (default: "./tts_output.wav")
-- `--language LANGUAGE`: Language for the TTS model, one of `'english_2026-01'`, `'english_2026-04'`, `'english'`, `'french_24l'`, `'german_24l'`, `'portuguese_24l'`, `'italian_24l'`, `'spanish_24l'` (default: `english`, which is the same model as `'english_2026-04'`). Incompatible with `--config`. The "24l" variants are bigger models, not distilled yet and here only as preview.
+- `--language LANGUAGE`: Language for the TTS model, one of `'english_2026-01'`, `'english_2026-04'`, `'english'`, `'french_24l'`, `'german'`, `'german_24l'`, `'portuguese'`, `'portuguese_24l'`, `'italian'`, `'italian_24l'`, `'spanish'`, `'spanish_24l'` (default: `english`, which is the same model as `'english_2026-04'`). Incompatible with `--config`. The "24l" variants are bigger models, not distilled yet and here only as preview.
 
 ### Generation Parameters
 
 - `--config CONFIG_PATH`: Path to custom config.yaml (for loading local model files). Incompatible with `--language`.
 - `--lsd-decode-steps LSD_DECODE_STEPS`: Number of generation steps (default: 1)
-- `--temperature TEMPERATURE`: Temperature for generation (default: 0.7)
+- `--temperature TEMPERATURE`: Temperature for generation (default: the model's recommended value from its config — 0.3 for the English model, 0.7 otherwise)
 - `--noise-clamp NOISE_CLAMP`: Noise clamp value (default: None)
 - `--eos-threshold EOS_THRESHOLD`: EOS threshold (default: -4.0)
 - `--frames-after-eos FRAMES_AFTER_EOS`: Number of frames to generate after EOS (default: None, auto-calculated based on the text length). Each frame is 80ms.

fastmcp_server.py

@@ -0,0 +1,127 @@
+import base64
+import io
+import json
+import os
+
+import numpy as np
+import scipy.io.wavfile as wavfile
+from fastmcp import FastMCP
+
+from pocket_tts import TTSModel
+from pocket_tts.default_parameters import get_default_voice_for_language
+from pocket_tts.utils.utils import _ORIGINS_OF_PREDEFINED_VOICES
+
+DEFAULT_LANGUAGE = os.environ.get("POCKET_TTS_MCP_LANGUAGE", "spanish_24l")
+DEFAULT_TEMP = float(os.environ.get("POCKET_TTS_MCP_TEMP", "0.5"))
+DEFAULT_LSD_DECODE_STEPS = int(os.environ.get("POCKET_TTS_MCP_LSD_STEPS", "5"))
+DEFAULT_EOS_THRESHOLD = float(os.environ.get("POCKET_TTS_MCP_EOS_THRESHOLD", "-3.0"))
+DEFAULT_HOST = os.environ.get("POCKET_TTS_MCP_HOST", "0.0.0.0")
+DEFAULT_PORT = int(os.environ.get("POCKET_TTS_MCP_PORT", "8008"))
+
+_model: TTSModel | None = None
+
+mcp = FastMCP(
+    name="Pocket TTS MCP",
+    instructions=(
+        "Generate WAV speech audio with Pocket TTS and inspect the built-in voice presets."
+    ),
+)
+
+
+def get_model() -> TTSModel:
+    global _model
+    if _model is None:
+        print("Loading Pocket TTS model for FastMCP server...")
+        _model = TTSModel.load_model(
+            language=DEFAULT_LANGUAGE,
+            temp=DEFAULT_TEMP,
+            lsd_decode_steps=DEFAULT_LSD_DECODE_STEPS,
+            eos_threshold=DEFAULT_EOS_THRESHOLD,
+        )
+    return _model
+
+
+def list_predefined_voices() -> list[dict[str, str | bool]]:
+    default_voice = get_default_voice_for_language(DEFAULT_LANGUAGE)
+    return [
+        {
+            "name": name,
+            "source": source,
+            "is_default": name == default_voice,
+        }
+        for name, source in sorted(_ORIGINS_OF_PREDEFINED_VOICES.items())
+    ]
+
+
+def encode_wav_base64(audio_tensor, sample_rate: int) -> tuple[str, int]:
+    if hasattr(audio_tensor, "detach"):
+        audio_numpy = audio_tensor.detach().cpu().numpy()
+    else:
+        audio_numpy = np.array(audio_tensor)
+
+    if audio_numpy.ndim > 1:
+        audio_numpy = audio_numpy.squeeze()
+
+    audio_numpy = np.asarray(audio_numpy, dtype=np.float32)
+
+    buffer = io.BytesIO()
+    wavfile.write(buffer, sample_rate, audio_numpy)
+    wav_bytes = buffer.getvalue()
+    return base64.b64encode(wav_bytes).decode("ascii"), len(wav_bytes)
+
+
+@mcp.tool()
+def list_voices() -> dict:
+    """List the built-in Pocket TTS voice presets available to the MCP client."""
+    return {
+        "language": DEFAULT_LANGUAGE,
+        "default_voice": get_default_voice_for_language(DEFAULT_LANGUAGE),
+        "voices": list_predefined_voices(),
+    }
+
+
+@mcp.resource("voice://predefined")
+def predefined_voices_resource() -> str:
+    """Return the predefined voice catalog as JSON."""
+    return json.dumps(list_voices(), indent=2)
+
+
+@mcp.tool()
+def generate_speech(
+    text: str,
+    voice: str | None = None,
+    frames_after_eos: int | None = None,
+    copy_state: bool = True,
+) -> dict:
+    """Generate speech audio and return a base64-encoded WAV payload."""
+    if not text.strip():
+        raise ValueError("text must not be empty")
+
+    model = get_model()
+    selected_voice = voice or get_default_voice_for_language(DEFAULT_LANGUAGE)
+    voice_state = model.get_state_for_audio_prompt(selected_voice)
+    audio_tensor = model.generate_audio(
+        model_state=voice_state,
+        text_to_generate=text,
+        frames_after_eos=frames_after_eos,
+        copy_state=copy_state,
+    )
+    audio_base64, wav_size_bytes = encode_wav_base64(audio_tensor, model.sample_rate)
+
+    return {
+        "text": text,
+        "voice": selected_voice,
+        "language": DEFAULT_LANGUAGE,
+        "sample_rate": model.sample_rate,
+        "mime_type": "audio/wav",
+        "wav_size_bytes": wav_size_bytes,
+        "audio_base64": audio_base64,
+    }
+
+
+def main() -> None:
+    mcp.run(transport="sse", host=DEFAULT_HOST, port=DEFAULT_PORT)
+
+
+if __name__ == "__main__":
+    main()

pocket_tts/config/english.yaml

@@ -2,6 +2,8 @@
 
 weights_path: hf://kyutai/pocket-tts/languages/english/model.safetensors@39592ff23c9ef80098bb74895d104c26275fe2c9
 weights_path_without_voice_cloning: hf://kyutai/pocket-tts-without-voice-cloning/languages/english/model.safetensors@d29db7978e464fb90cb3359ee0c69a273b9142cc
+# Human evals preferred this model at 0.3 (equal WER/similarity/speech rate).
+default_temperature: 0.3
 
 
 flow_lm:

pocket_tts/config/english_2026-04.yaml

@@ -2,6 +2,8 @@
 
 weights_path: hf://kyutai/pocket-tts/languages/english_2026-04/model.safetensors@19f95fe2df36e79fbd9f10008595cc4c977a0fcc
 weights_path_without_voice_cloning: hf://kyutai/pocket-tts-without-voice-cloning/languages/english_2026-04/model.safetensors@d29db7978e464fb90cb3359ee0c69a273b9142cc
+# Human evals preferred this model at 0.3 (equal WER/similarity/speech rate).
+default_temperature: 0.3
 
 
 flow_lm:

pocket_tts/main.py

@@ -20,7 +20,6 @@ from pocket_tts.default_parameters import (
     DEFAULT_FRAMES_AFTER_EOS,
     DEFAULT_LSD_DECODE_STEPS,
     DEFAULT_NOISE_CLAMP,
-    DEFAULT_TEMPERATURE,
     MAX_TOKEN_PER_CHUNK,
     get_default_text_for_language,
     get_default_voice_for_language,
@@ -261,8 +260,12 @@ def generate(
         int, typer.Option(help="Number of generation steps")
     ] = DEFAULT_LSD_DECODE_STEPS,
     temperature: Annotated[
-        float, typer.Option(help="Temperature for generation")
-    ] = DEFAULT_TEMPERATURE,
+        float | None,
+        typer.Option(
+            help="Temperature for generation. Defaults to the model's recommended "
+            "value from its config (0.3 for the English model, 0.7 otherwise)."
+        ),
+    ] = None,
     noise_clamp: Annotated[float, typer.Option(help="Noise clamp value")] = DEFAULT_NOISE_CLAMP,
     eos_threshold: Annotated[float, typer.Option(help="EOS threshold")] = DEFAULT_EOS_THRESHOLD,
     frames_after_eos: Annotated[

pocket_tts/models/tts_model.py

@@ -25,7 +25,6 @@ from pocket_tts.default_parameters import (
     DEFAULT_LANGUAGE,
     DEFAULT_LSD_DECODE_STEPS,
     DEFAULT_NOISE_CLAMP,
-    DEFAULT_TEMPERATURE,
     MAX_TOKEN_PER_CHUNK,
 )
 from pocket_tts.models.flow_lm import FlowLMModel
@@ -234,7 +233,7 @@ class TTSModel(nn.Module):
         cls,
         language: str | None = None,
         config: str | Path | None = None,
-        temp: float | int = DEFAULT_TEMPERATURE,
+        temp: float | int | None = None,
         lsd_decode_steps: int = DEFAULT_LSD_DECODE_STEPS,
         noise_clamp: float | int | None = DEFAULT_NOISE_CLAMP,
         eos_threshold: float = DEFAULT_EOS_THRESHOLD,
@@ -253,7 +252,9 @@ class TTSModel(nn.Module):
                 If neither `config` nor `language` is provided, defaults to `"english", which is the same model as 'english_2026-04'`.
             config: A path to a custom YAML config file saved locally (e.g., `"C://pocket_tts/pocket_tts_config.yaml"`).
             temp: Sampling temperature for generation. Higher values produce more
-                diverse but potentially lower quality output.
+                diverse but potentially lower quality output. If None, defaults to
+                the model's recommended value from its config file
+                (``default_temperature``, e.g. 0.3 for the English model).
             lsd_decode_steps: Number of steps for Lagrangian Self Distillation
                 decoding. More steps can improve quality but increase computation.
             noise_clamp: Maximum value for noise sampling. If None, no clamping
@@ -303,6 +304,8 @@ class TTSModel(nn.Module):
             raise ValueError("Config should be a path to a YAML file ending with .yaml")
         config_path = Path(config)
         config = load_config(config_path)
+        if temp is None:
+            temp = config.default_temperature
         logger.info(f"Loading model from config at {config_path}...")
 
         tts_model = TTSModel._from_pydantic_config_with_weights(

pocket_tts/modules/transformer.py

@@ -118,8 +118,15 @@ class StreamingMultiheadAttention(StatefulModule):
         self.out_proj = nn.Linear(embed_dim, mult * embed_dim, bias=False)
 
     def init_state(self, batch_size: int, sequence_length: int) -> dict[str, torch.Tensor]:
-        device = self.in_proj.weight.device
-        dtype = self.in_proj.weight.dtype
+        weight = self.in_proj.weight
+        if callable(weight):
+            # torch.ao dynamic-quantized Linear: weight is a method returning a
+            # qint8 tensor, but activations stay float32 — that's what the cache holds.
+            device = weight().device
+            dtype = torch.float32
+        else:
+            device = weight.device
+            dtype = weight.dtype
         return self._cache_backend.init_state(batch_size, sequence_length, device, dtype)
 
     def increment_step(self, state: dict, increment: int = 1):

pocket_tts/utils/config.py

@@ -116,6 +116,7 @@ class Config(StrictModel):
     pad_with_spaces_for_short_inputs: bool = False
     remove_semicolons: bool = False
     model_recommended_frames_after_eos: int | None = None
+    default_temperature: float = 0.7
 
 
 def load_config(yaml_path: str | Path) -> Config:

pyproject.toml

@@ -2,7 +2,7 @@
 name = "pocket-tts"
 requires-python = ">= 3.10,<3.15"
 description = "Kyutai's pocket-sized text-to-speech!"
-version = "2.0.0"
+version = "2.1.0"
 readme = "README.md"
 dependencies = [
     "numpy>=2",
@@ -18,9 +18,11 @@ dependencies = [
     "uvicorn>=0.13.0",
     "python-multipart>=0.0.21",
     "scipy>=1.5.0",
+    "soundfile>=0.12.0",
     "einops>=0.4.0",
     "huggingface_hub>=0.13.0",
     "requests>=2.20.0",
+    "fastmcp>=2.0.0",
 ]
 
 [project.optional-dependencies]
@@ -70,6 +72,7 @@ packages = ["pocket_tts"]
 
 [project.scripts]
 pocket-tts = "pocket_tts.main:cli_app"
+pocket-tts-mcp = "fastmcp_server:main"
 
 [tool.uv]
 # System python distribution can have header files missing, so we use only the python

server.py

@@ -0,0 +1,288 @@
+import io
+import re
+import traceback
+from typing import Literal
+
+import numpy as np
+import scipy.io.wavfile as wavfile
+from fastapi import FastAPI, Form, HTTPException
+from fastapi.openapi.docs import get_swagger_ui_html
+from fastapi.responses import RedirectResponse, Response
+from pocket_tts import TTSModel
+from pocket_tts.models.tts_model import prepare_text_prompt, split_into_best_sentences
+
+app = FastAPI(
+    title="Pocket TTS High-Quality API",
+    description="HTTP API for generating MP3 or WAV speech audio with Pocket TTS.",
+    version="1.0.0",
+    docs_url=None,
+    redoc_url="/redoc",
+    openapi_url="/openapi.json",
+)
+
+SWAGGER_UI_VERSION = "5.32.10"
+SWAGGER_UI_JS_URL = (
+    f"https://cdn.jsdelivr.net/npm/swagger-ui-dist@{SWAGGER_UI_VERSION}/swagger-ui-bundle.js"
+)
+SWAGGER_UI_CSS_URL = (
+    f"https://cdn.jsdelivr.net/npm/swagger-ui-dist@{SWAGGER_UI_VERSION}/swagger-ui.css"
+)
+SAFE_MAX_TOKENS_PER_CHUNK = 35
+JOIN_SILENCE_MS = 100
+MIN_TOKENS_PER_CHUNK = 12
+
+print("Caricamento del modello Pocket TTS in memoria...")
+model = TTSModel.load_model(
+    language="spanish_24l", temp=0.5, lsd_decode_steps=5, eos_threshold=-3.0
+)
+last_generated_audio: dict[str, bytes | str] | None = None
+
+
+@app.get("/", include_in_schema=False)
+async def swagger_ui_redirect():
+    return RedirectResponse(url="/docs")
+
+
+@app.get("/docs", include_in_schema=False)
+async def custom_swagger_ui():
+    return get_swagger_ui_html(
+        openapi_url=app.openapi_url,
+        title=f"{app.title} - Swagger UI",
+        swagger_js_url=SWAGGER_UI_JS_URL,
+        swagger_css_url=SWAGGER_UI_CSS_URL,
+        swagger_ui_parameters={
+            "displayRequestDuration": True,
+        },
+    )
+
+
+def build_audio_response(audio_bytes: bytes, media_type: str) -> Response:
+    return Response(
+        content=audio_bytes,
+        media_type=media_type,
+        headers={"Cache-Control": "no-store"},
+    )
+
+
+def normalize_tts_text(text: str) -> str:
+    def decode_unicode_match(match: re.Match[str]) -> str:
+        return match.group(0).encode("ascii").decode("unicode_escape")
+
+    if "\\u" in text:
+        text = re.sub(r"\\u[0-9a-fA-F]{4}", decode_unicode_match, text)
+
+    text = text.replace("\\n", "\n").replace("\\t", " ")
+
+    # Keep heading content while removing markdown wrappers.
+    text = re.sub(r"\*\*\s*\[([^\]]+)\]\s*\*\*", r"\n\1.\n", text)
+    text = re.sub(r"\[([^\]]+)\]", r"\1", text)
+    text = text.replace("**", "").replace("__", "")
+
+    # Remove list/heading markdown markers that harm pronunciation.
+    text = re.sub(r"^[#>*-]+\s*", "", text, flags=re.MULTILINE)
+
+    # Normalize whitespace while preserving sentence boundaries.
+    text = re.sub(r"\r\n?", "\n", text)
+    text = re.sub(r"\n{2,}", "\n\n", text)
+    text = re.sub(r"[ \t]+", " ", text)
+    text = re.sub(r" *\n *", "\n", text)
+
+    return text.strip()
+
+
+def tensor_to_numpy(audio_tensor) -> np.ndarray:
+    if hasattr(audio_tensor, "detach"):
+        audio_numpy = audio_tensor.detach().cpu().numpy()
+    else:
+        audio_numpy = np.array(audio_tensor)
+
+    if audio_numpy.ndim > 1:
+        audio_numpy = audio_numpy.squeeze()
+
+    return np.asarray(audio_numpy, dtype=np.float32)
+
+
+def count_model_tokens(text: str) -> int:
+    tokenizer = model.flow_lm.conditioner.tokenizer
+    prepared_text, _ = prepare_text_prompt(
+        text,
+        model.pad_with_spaces_for_short_inputs,
+        model.remove_semicolons,
+    )
+    return len(tokenizer(prepared_text.strip()).tokens[0].tolist())
+
+
+def split_text_into_safe_chunks(text: str, max_tokens: int) -> list[str]:
+    tokenizer = model.flow_lm.conditioner.tokenizer
+    chunks = split_into_best_sentences(
+        tokenizer,
+        text,
+        max_tokens,
+        model.pad_with_spaces_for_short_inputs,
+        remove_semicolons=model.remove_semicolons,
+    )
+
+    refined_chunks: list[str] = []
+    for chunk in chunks:
+        chunk = chunk.strip()
+        token_count = count_model_tokens(chunk)
+        if token_count <= max_tokens:
+            refined_chunks.append(chunk)
+            continue
+
+        words = chunk.split()
+        current_words: list[str] = []
+        for word in words:
+            candidate_words = current_words + [word]
+            candidate_text = " ".join(candidate_words)
+            candidate_tokens = count_model_tokens(candidate_text)
+            if current_words and candidate_tokens > max_tokens:
+                refined_chunks.append(" ".join(current_words))
+                current_words = [word]
+            else:
+                current_words = candidate_words
+
+        if current_words:
+            refined_chunks.append(" ".join(current_words))
+
+    merged_chunks: list[str] = []
+    for chunk in refined_chunks:
+        chunk = chunk.strip()
+        if not chunk:
+            continue
+
+        chunk_tokens = count_model_tokens(chunk)
+        if not merged_chunks:
+            merged_chunks.append(chunk)
+            continue
+
+        previous_chunk = merged_chunks[-1]
+        previous_tokens = count_model_tokens(previous_chunk)
+        combined = f"{previous_chunk} {chunk}".strip()
+        combined_tokens = count_model_tokens(combined)
+
+        # Avoid tiny fragments when the merge still fits comfortably.
+        if chunk_tokens < MIN_TOKENS_PER_CHUNK and combined_tokens <= max_tokens:
+            merged_chunks[-1] = combined
+            continue
+
+        # Also merge if the previous chunk is tiny and the combined text still fits.
+        if previous_tokens < MIN_TOKENS_PER_CHUNK and combined_tokens <= max_tokens:
+            merged_chunks[-1] = combined
+            continue
+
+        merged_chunks.append(chunk)
+
+    for chunk in merged_chunks:
+        token_count = count_model_tokens(chunk)
+        if token_count > max_tokens:
+            raise ValueError(
+                f"Chunk exceeds hard token limit ({token_count} > {max_tokens}): {chunk[:120]!r}"
+            )
+
+    return merged_chunks
+
+
+def generate_audio_from_chunks(voice_state: dict, text: str, sample_rate: int) -> np.ndarray:
+    chunks = split_text_into_safe_chunks(text, SAFE_MAX_TOKENS_PER_CHUNK)
+    audio_segments: list[np.ndarray] = []
+    silence = np.zeros(int(sample_rate * JOIN_SILENCE_MS / 1000), dtype=np.float32)
+
+    for index, chunk in enumerate(chunks):
+        audio_tensor = model.generate_audio(
+            voice_state,
+            chunk,
+            max_tokens=SAFE_MAX_TOKENS_PER_CHUNK,
+            copy_state=True,
+        )
+        audio_segments.append(tensor_to_numpy(audio_tensor))
+        if index < len(chunks) - 1:
+            audio_segments.append(silence.copy())
+
+    if not audio_segments:
+        return np.zeros(0, dtype=np.float32)
+
+    return np.concatenate(audio_segments)
+
+
+@app.get("/v1/audio/speech", include_in_schema=False)
+@app.get("/tts", include_in_schema=False)
+async def get_last_generated_audio():
+    if last_generated_audio is None:
+        raise HTTPException(status_code=404, detail="No generated audio available yet")
+    return build_audio_response(
+        audio_bytes=last_generated_audio["audio_bytes"],
+        media_type=last_generated_audio["media_type"],
+    )
+
+
+@app.post("/v1/audio/speech")
+@app.post(
+    "/tts",
+    tags=["tts"],
+    summary="Generate speech audio",
+    description=(
+        "Generate speech from text and return the result as an MP3 file by default, "
+        "or as WAV when `format=wav`. "
+        "The same handler is available on `/v1/audio/speech` and `/tts`."
+    ),
+    response_class=Response,
+    responses={
+        200: {
+            "description": "Generated MP3 or WAV audio",
+            "content": {"audio/mpeg": {}, "audio/wav": {}},
+        },
+        500: {"description": "Audio generation failed"},
+    },
+)
+async def generate_tts(
+    text: str = Form(...),
+    voice_url: str = Form("lola"),
+    format: Literal["mp3", "wav"] = Form("mp3"),
+):
+    global last_generated_audio
+    try:
+        normalized_text = normalize_tts_text(text)
+        # 1. Recupera lo stato del prompt vocale (dict)
+        voice_state = model.get_state_for_audio_prompt(voice_url)
+        sample_rate = model.sample_rate if hasattr(model, "sample_rate") else 24000
+        audio_numpy = generate_audio_from_chunks(voice_state, normalized_text, sample_rate)
+
+        buffer = io.BytesIO()
+
+        if format == "wav":
+            wavfile.write(buffer, sample_rate, audio_numpy)
+            media_type = "audio/wav"
+        else:
+            try:
+                import soundfile as sf
+            except ImportError as exc:
+                raise HTTPException(
+                    status_code=500,
+                    detail=(
+                        "MP3 output requires the optional `soundfile` dependency. "
+                        "Install it with `uv add soundfile`."
+                    ),
+                ) from exc
+
+            sf.write(buffer, audio_numpy, sample_rate, format="MP3")
+            media_type = "audio/mpeg"
+
+        buffer.seek(0)
+        audio_bytes = buffer.read()
+
+        last_generated_audio = {
+            "audio_bytes": audio_bytes,
+            "media_type": media_type,
+        }
+        return build_audio_response(audio_bytes=audio_bytes, media_type=media_type)
+
+    except Exception as e:
+        print("\n--- ERRORE DURANTE LA GENERAZIONE ---")
+        traceback.print_exc()
+        print("-------------------------------------\n")
+        raise HTTPException(status_code=500, detail=str(e))
+
+if __name__ == "__main__":
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=8009)

start-pocke-tts.sh

@@ -0,0 +1,2 @@
+uv run --with "fastapi,uvicorn,scipy" python3 server.py
+