app.py

import os
import re
import io
import json
import time
import base64
import asyncio
from typing import AsyncGenerator, List

import numpy as np
from fastapi import FastAPI, Query
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import StreamingResponse, JSONResponse

from kokoro_onnx import Kokoro  # pip install kokoro-onnx

APP_NAME = "kokoro-onnx-fastapi-sse"
SAMPLE_RATE = 24000            # Kokoro output
CHANNELS = 1
CHUNK_SAMPLES = 2400           # 100 ms

MODEL_PATH = os.getenv("KOKORO_MODEL", "models/kokoro-v1.0.int8.onnx")
VOICES_PATH = os.getenv("KOKORO_VOICES", "models/voices-v1.0.bin")

app = FastAPI(title=APP_NAME)
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"], allow_methods=["*"], allow_headers=["*"]
)

kokoro = None
_model_lock = asyncio.Lock()


def split_text(text: str, max_len: int = 220) -> List[str]:
    # Split on sentences, then fold long pieces
    parts = re.split(r"(?<=[\.\!\?।])\s+", text.strip())
    chunks: List[str] = []
    for p in parts:
        p = p.strip()
        while len(p) > max_len:
            cut = p.rfind(" ", 0, max_len)
            if cut == -1:
                cut = max_len
            chunks.append(p[:cut].strip())
            p = p[cut:].strip()
        if p:
            chunks.append(p)
    return [c for c in chunks if c]


def floats_to_s16le(samples: np.ndarray) -> np.ndarray:
    x = np.clip(samples, -1.0, 1.0)
    return (x * 32767.0).astype(np.int16)


async def sse_generator(text: str, voice: str, speed: float, lang: str) -> AsyncGenerator[bytes, None]:
    # Yields SSE messages with base64 PCM16 frames (100ms per chunk).
    seq = 0
    total_samples = 0
    try:
        # Warmup ping
        yield b": keep-alive\n\n"
        for sentence in split_text(text):
            async with _model_lock:
                samples, sr = kokoro.create(sentence, voice=voice, speed=speed, lang=lang)
            assert sr == SAMPLE_RATE, f"Expected {SAMPLE_RATE}, got {sr}"
            pcm16 = floats_to_s16le(np.asarray(samples))

            for i in range(0, len(pcm16), CHUNK_SAMPLES):
                frame = pcm16[i:i+CHUNK_SAMPLES]
                total_samples += len(frame)
                payload = {
                    "seq": seq,
                    "sr": SAMPLE_RATE,
                    "ch": CHANNELS,
                    "format": "s16le",
                    "pcm16": base64.b64encode(frame.tobytes()).decode("ascii"),
                }
                msg = f"data: {json.dumps(payload, separators=(',',':'))}\n\n"
                yield msg.encode("utf-8")
                seq += 1
                await asyncio.sleep(0)  # give control back to loop
        # Done event
        done = {"total_chunks": seq, "total_samples": total_samples}
        yield f"event: done\ndata: {json.dumps(done)}\n\n".encode("utf-8")
    except asyncio.CancelledError:
        # Client disconnected
        return


@app.on_event("startup")
async def _load_model():
    global kokoro
    kokoro = Kokoro(MODEL_PATH, VOICES_PATH)


@app.get("/healthz")
async def healthz():
    return {"status": "ok", "model": os.path.basename(MODEL_PATH)}


@app.get("/v1/voices")
async def list_voices():
    try:
        import numpy as _np
        with _np.load(VOICES_PATH) as z:
            names = sorted(list(z.files))
        return {"voices": names}
    except Exception:
        fallback = [
            "af", "af_bella", "af_nicole", "af_sarah", "af_sky",
            "am_adam", "am_michael",
            "bf_emma", "bf_isabella",
            "bm_george", "bm_lewis",
        ]
        return {"voices": fallback, "note": "fallback list; voices file not parsed"}


@app.get("/v1/tts.sse")
async def tts_sse(
    text: str = Query(..., description="Text to synthesize"),
    voice: str = Query("af_sarah"),
    speed: float = Query(1.0, ge=0.5, le=1.5),
    lang: str = Query("en-us"),
):
    headers = {
        "Cache-Control": "no-cache",
        "Connection": "keep-alive",
        "X-Accel-Buffering": "no",  # for nginx
    }
    return StreamingResponse(
        sse_generator(text, voice, speed, lang),
        media_type="text/event-stream",
        headers=headers,
    )