An OpenAI-compatible TTS server on your own computer

Install

You need Python 3.10 or newer. Install the engine, the web server and the audio encoder:

pip install superclonica-cpu fastapi uvicorn av

On the first run the engine downloads the Supertonic 3 model, about 400 MB; after that it works offline on an ordinary processor, and an NVIDIA graphics card only makes it faster.

The server

Save this as openai_tts.py:

import io, os, wave
import av
import numpy as np
from fastapi import FastAPI, Header, HTTPException
from fastapi.concurrency import run_in_threadpool
from fastapi.responses import Response
from pydantic import BaseModel
from superclonica import Superclonica

VOICES = {name: name for name in ["M1", "M2", "M3", "M4", "M5", "F1", "F2", "F3", "F4", "F5"]}
# a catalog voice: VOICES["narrator"] = "/path/to/voice-file-from-the-catalog"
LANGUAGE = os.environ.get("TTS_LANGUAGE", "en")
KEY = os.environ.get("TTS_KEY", "")  # leave empty on a private machine
app = FastAPI()
tts = Superclonica()

class Speech(BaseModel):
    input: str
    voice: str = "F1"
    model: str = "superclonica"
    response_format: str = "mp3"

def to_wav(pcm, rate):
    buffer = io.BytesIO()
    with wave.open(buffer, "wb") as file:
        file.setnchannels(1)
        file.setsampwidth(2)
        file.setframerate(rate)
        file.writeframes(pcm.tobytes())
    return buffer.getvalue()

def to_mp3(pcm, rate):
    buffer = io.BytesIO()
    with av.open(buffer, "w", format="mp3") as container:
        stream = container.add_stream("libmp3lame", rate=rate, layout="mono")
        frame = av.AudioFrame.from_ndarray(pcm.reshape(1, -1), format="s16", layout="mono")
        frame.rate = rate
        for packet in stream.encode(frame):
            container.mux(packet)
        for packet in stream.encode(None):
            container.mux(packet)
    return buffer.getvalue()

def speak(request):
    voice = VOICES.get(request.voice, "F1")
    audio, rate = tts.synthesize(request.input, voice=voice, lang=LANGUAGE)
    pcm = (np.clip(np.asarray(audio, dtype=np.float32).reshape(-1), -1, 1) * 32767).astype(np.int16)
    if request.response_format == "wav":
        return to_wav(pcm, rate), "audio/wav"
    if request.response_format == "pcm":
        return pcm.tobytes(), "audio/pcm"
    return to_mp3(pcm, rate), "audio/mpeg"

@app.post("/v1/audio/speech")
async def speech(request: Speech, authorization: str = Header("")):
    if KEY and authorization != "Bearer " + KEY:
        raise HTTPException(401)
    audio, media_type = await run_in_threadpool(speak, request)
    return Response(audio, media_type=media_type)

@app.get("/v1/models")
def models():
    return {"object": "list", "data": [{"id": "superclonica", "object": "model"}]}

@app.get("/v1/audio/voices")
def voices():
    return {"voices": list(VOICES)}

Start it, and it answers on port 8880:

uvicorn openai_tts:app --host 127.0.0.1 --port 8880

The voice field takes one of the ten stock voices, M1 to M5 and F1 to F5; unknown names fall back to F1, so a client that sends alloy still gets an answer. Set TTS_LANGUAGE to the language your app talks in: the engine speaks 31 languages, but not Chinese yet. Set TTS_KEY if anyone but you can reach the port.

Connect Open WebUI

Open Admin Panel, Settings, Audio. Set the Text-to-Speech Engine to OpenAI, the API Base URL to http://localhost:8880/v1, the TTS Model to superclonica and the TTS Voice to one of the stock voices. Open WebUI wants a non-empty API key: type your TTS_KEY, or any text if you left it empty. If Open WebUI runs in Docker, start the server with --host 0.0.0.0 and use http://host.docker.internal:8880/v1; on Linux the container needs --add-host=host.docker.internal:host-gateway. Leave Response Splitting on punctuation so long answers start speaking after the first sentence.

Connect SillyTavern

Open Extensions, TTS, and choose the OpenAI Compatible provider. Enter http://localhost:8880/v1/audio/speech as the endpoint, superclonica as the model, your key, and type the voice list yourself, because this provider does not fetch it: M1,M2,M3,M4,M5,F1,F2,F3,F4,F5. Then give every character and your persona their own voice in the voice map and click Apply.

Character voices and other clients

To use a voice from the SuperClonica catalog, download its voice file and add a line such as VOICES["narrator"] = "/path/to/the/file"; the name narrator then works in every client, including the voice list in SillyTavern. Any program built on the OpenAI SDK works too: create the client with base_url set to http://localhost:8880/v1 and call audio.speech.create as usual.