scripts/04_streaming_server.py

#!/usr/bin/env python3
"""
================================================================================
Priority 4: Sentence-Level Streaming TTS Server with Opus Output
================================================================================

Production TTS requires streaming audio, not waiting for full generation.
This server implements:

1. Sentence-level chunking: Split input at punctuation boundaries
2. Per-sentence generation: Generate mel-spectrogram for each sentence
3. Opus encoding: Stream compressed audio chunks as each sentence completes
4. Time-to-first-audio (TTFA): Sub-500ms for short sentences

Architecture:
    Client → HTTP POST /synthesize
         → Preprocess (diacritize, normalize)
         → Chunk into sentences
         → For each sentence:
             → Generate with EPSS(7) + BF16
             → Encode to Opus
             → Yield audio chunk
         → Client receives streaming audio

Opus encoding:
    - 24kHz sample rate (matches F5-TTS output)
    - 16-32kbps bitrate
    - ~10x smaller than WAV
    - Streaming-compatible (Ogg Opus container)

Dependencies:
    pip install fastapi uvicorn opuslib pydub

Usage:
    # Start server
    python 04_streaming_server.py --host 0.0.0.0 --port 8000

    # Client request
    curl -X POST http://localhost:8000/synthesize \
        -H "Content-Type: application/json" \
        -d '{
            "text": "مرحبا بك. كيف حالك اليوم؟",
            "ref_audio": "reference.wav",
            "ref_text": "مرحبا"
        }' \
        --output output.opus

================================================================================
"""

import argparse
import asyncio
import io
import os
import sys
import time
import warnings
from contextlib import asynccontextmanager
from pathlib import Path
from typing import AsyncGenerator, List, Optional

import numpy as np
import soundfile as sf
import torch
import torchaudio
from cached_path import cached_path
from fastapi import FastAPI, HTTPException, Request
from fastapi.responses import StreamingResponse
from f5_tts.infer.utils_infer import load_vocoder, preprocess_ref_audio_text
from f5_tts.model import CFM
from f5_tts.model.utils import get_tokenizer
from habibi_tts.model.utils import dialect_id_map, text_list_formatter
from hydra.utils import get_class
from omegaconf import OmegaConf
from pydantic import BaseModel

warnings.filterwarnings("ignore")

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
MODEL_CFG_PATH = str(Path(__file__).parent / "configs" / "F5TTS_v1_Base.yaml")
CKPT_URL = "hf://SWivid/Habibi-TTS/Specialized/ALG/model_100000.safetensors"
VOCAB_URL = "hf://SWivid/Habibi-TTS/Specialized/ALG/vocab.txt"

N_MEL_CHANNELS = 100
HOP_LENGTH = 256
WIN_LENGTH = 1024
N_FFT = 1024
TARGET_SAMPLE_RATE = 24000

# ---------------------------------------------------------------------------
# Global model state (loaded once at startup)
# ---------------------------------------------------------------------------

model_global = None
vocoder_global = None


# ---------------------------------------------------------------------------
# Model Loading
# ---------------------------------------------------------------------------


def load_production_model(device=DEVICE):
    """Load optimized Habibi-TTS ALG model for production."""
    print(f"[LOAD] Loading production model on {device}...")

    model_cfg = OmegaConf.load(MODEL_CFG_PATH)
    model_cls = get_class(f"f5_tts.model.{model_cfg.model.backbone}")
    model_arc = model_cfg.model.arch

    ckpt_file = str(cached_path(CKPT_URL))
    vocab_file = str(cached_path(VOCAB_URL))

    vocab_char_map, vocab_size = get_tokenizer(vocab_file, "custom")

    model = CFM(
        transformer=model_cls(**model_arc, text_num_embeds=vocab_size, mel_dim=N_MEL_CHANNELS),
        mel_spec_kwargs=dict(
            n_fft=N_FFT,
            hop_length=HOP_LENGTH,
            win_length=WIN_LENGTH,
            n_mel_channels=N_MEL_CHANNELS,
            target_sample_rate=TARGET_SAMPLE_RATE,
            mel_spec_type="vocos",
        ),
        odeint_kwargs=dict(method="euler"),
        vocab_char_map=vocab_char_map,
    ).to(device)

    # Load checkpoint
    from safetensors.torch import load_file
    checkpoint = load_file(ckpt_file, device=device)
    checkpoint = {"ema_model_state_dict": checkpoint}
    checkpoint["model_state_dict"] = {
        k.replace("ema_model.", ""): v
        for k, v in checkpoint["ema_model_state_dict"].items()
        if k not in ["initted", "step"]
    }
    for key in ["mel_spec.mel_stft.mel_scale.fb", "mel_spec.mel_stft.spectrogram.window"]:
        if key in checkpoint["model_state_dict"]:
            del checkpoint["model_state_dict"][key]
    model.load_state_dict(checkpoint["model_state_dict"])
    del checkpoint
    torch.cuda.empty_cache()

    # BF16 optimization
    if device == "cuda":
        model = model.to(torch.bfloat16)
        print("[OPT] Model converted to BF16")

    # torch.compile for transformer backbone
    if device == "cuda":
        model.transformer = torch.compile(model.transformer, mode="reduce-overhead", fullgraph=False)
        print("[OPT] torch.compile applied")

    model.eval()
    return model


# ---------------------------------------------------------------------------
# Audio Processing
# ---------------------------------------------------------------------------


def chunk_text(text: str, max_chars: int = 135) -> List[str]:
    """Split text into sentence-level chunks."""
    import re
    sentences = re.split(r"(?<=[;:,.!?])\s+|(?<=[؛：،。！？])", text)
    chunks = []
    current_chunk = ""

    for sentence in sentences:
        if not sentence.strip():
            continue
        if len(current_chunk.encode("utf-8")) + len(sentence.encode("utf-8")) <= max_chars:
            current_chunk += sentence + " " if sentence and sentence[-1].isascii() else sentence
        else:
            if current_chunk:
                chunks.append(current_chunk.strip())
            current_chunk = sentence + " " if sentence and sentence[-1].isascii() else sentence

    if current_chunk:
        chunks.append(current_chunk.strip())

    return chunks


def wav_to_opus_bytes(wav: np.ndarray, sr: int = 24000, bitrate: str = "24k") -> bytes:
    """Convert WAV numpy array to Opus-encoded bytes."""
    try:
        import subprocess
        # Write to temporary WAV
        wav_int16 = (wav * 32767).astype(np.int16)
        wav_buffer = io.BytesIO()
        sf.write(wav_buffer, wav_int16, sr, format="WAV", subtype="PCM_16")
        wav_bytes = wav_buffer.getvalue()

        # Convert to Opus using ffmpeg
        proc = subprocess.run(
            ["ffmpeg", "-i", "-", "-c:a", "libopus", "-b:a", bitrate, "-f", "ogg", "-"],
            input=wav_bytes,
            capture_output=True,
        )
        if proc.returncode != 0:
            # Fallback: return WAV if opus encoding fails
            return wav_bytes
        return proc.stdout
    except Exception:
        # Fallback to WAV
        wav_buffer = io.BytesIO()
        sf.write(wav_buffer, wav, sr, format="WAV")
        return wav_buffer.getvalue()


# ---------------------------------------------------------------------------
# Streaming Inference
# ---------------------------------------------------------------------------


def infer_sentence(
    ref_audio: torch.Tensor,
    ref_text: str,
    gen_text: str,
    model_obj,
    vocoder,
    nfe_step: int = 7,
    cfg_strength: float = 2.0,
    sway_sampling_coef: float = -1.0,
    speed: float = 1.0,
    device: str = DEVICE,
) -> np.ndarray:
    """Generate audio for a single sentence."""
    audio = ref_audio.to(device)
    ref_audio_len = audio.shape[-1] // HOP_LENGTH

    # Prepare text with dialect ID
    text_list = [ref_text + gen_text]
    final_text_list = text_list_formatter(text_list, dialect_id=dialect_id_map["ALG"])

    # Calculate duration
    ref_text_len = len(ref_text.encode("utf-8"))
    gen_text_len = len(gen_text.encode("utf-8"))
    duration = ref_audio_len + int(ref_audio_len / ref_text_len * gen_text_len / speed)

    with torch.inference_mode():
        generated, _ = model_obj.sample(
            cond=audio,
            text=final_text_list,
            duration=duration,
            steps=nfe_step,
            cfg_strength=cfg_strength,
            sway_sampling_coef=sway_sampling_coef,
        )

        generated = generated.to(torch.float32)
        generated = generated[:, ref_audio_len:, :]
        generated = generated.permute(0, 2, 1)
        generated_wave = vocoder.decode(generated)
        generated_wave = generated_wave.squeeze().cpu().numpy()

    return generated_wave


async def stream_synthesize(
    text: str,
    ref_audio_path: str,
    ref_text: str,
    model_obj,
    vocoder,
    nfe_step: int = 7,
    device: str = DEVICE,
) -> AsyncGenerator[bytes, None]:
    """Stream synthesized audio in sentence-level chunks."""
    # Preprocess reference
    ref_audio_path, ref_text = preprocess_ref_audio_text(ref_audio_path, ref_text)

    # Load reference audio
    audio, sr = torchaudio.load(ref_audio_path)
    if audio.shape[0] > 1:
        audio = torch.mean(audio, dim=0, keepdim=True)
    if sr != TARGET_SAMPLE_RATE:
        resampler = torchaudio.transforms.Resample(sr, TARGET_SAMPLE_RATE)
        audio = resampler(audio)

    # Normalize RMS
    rms = torch.sqrt(torch.mean(torch.square(audio)))
    target_rms = 0.1
    if rms < target_rms:
        audio = audio * target_rms / rms

    # Chunk text
    sentences = chunk_text(text)
    print(f"[STREAM] Text split into {len(sentences)} chunks")

    # Generate and stream each sentence
    for i, sentence in enumerate(sentences):
        t0 = time.perf_counter()
        wav = infer_sentence(
            audio, ref_text, sentence, model_obj, vocoder,
            nfe_step=nfe_step, device=device,
        )
        t1 = time.perf_counter()

        # Re-normalize if needed
        if rms < target_rms:
            wav = wav * rms.item() / target_rms

        # Encode to Opus
        opus_bytes = wav_to_opus_bytes(wav, sr=TARGET_SAMPLE_RATE, bitrate="24k")

        print(f"[STREAM] Chunk {i+1}/{len(sentences)}: {len(sentence)} chars, "
              f"gen={t1-t0:.3f}s, audio={len(wav)/TARGET_SAMPLE_RATE:.2f}s")

        yield opus_bytes


# ---------------------------------------------------------------------------
# FastAPI Application
# ---------------------------------------------------------------------------


class SynthesizeRequest(BaseModel):
    text: str
    ref_audio: str
    ref_text: str = ""
    nfe_step: int = 7
    cfg_strength: float = 2.0
    sway_sampling_coef: float = -1.0
    speed: float = 1.0
    output_format: str = "opus"  # opus, wav


@asynccontextmanager
async def lifespan(app: FastAPI):
    """Load model at startup."""
    global model_global, vocoder_global
    print("[STARTUP] Loading production model...")
    model_global = load_production_model(device=DEVICE)
    vocoder_global = load_vocoder("vocos", is_local=False, local_path="", device=DEVICE)
    print("[STARTUP] Model loaded. Server ready.")
    yield
    print("[SHUTDOWN] Cleaning up...")
    del model_global, vocoder_global
    torch.cuda.empty_cache() if DEVICE == "cuda" else None


app = FastAPI(title="Habibi-TTS ALG Streaming Server", lifespan=lifespan)


@app.post("/synthesize")
async def synthesize(request: SynthesizeRequest):
    """Stream synthesized audio for the given text."""
    if not os.path.exists(request.ref_audio):
        raise HTTPException(status_code=400, detail=f"Reference audio not found: {request.ref_audio}")

    async def generate():
        async for chunk in stream_synthesize(
            request.text,
            request.ref_audio,
            request.ref_text,
            model_global,
            vocoder_global,
            nfe_step=request.nfe_step,
            device=DEVICE,
        ):
            yield chunk

    media_type = "audio/ogg" if request.output_format == "opus" else "audio/wav"
    return StreamingResponse(generate(), media_type=media_type)


@app.post("/synthesize_sync")
async def synthesize_sync(request: SynthesizeRequest):
    """Synchronous synthesis (full audio returned)."""
    if not os.path.exists(request.ref_audio):
        raise HTTPException(status_code=400, detail=f"Reference audio not found: {request.ref_audio}")

    chunks = []
    async for chunk in stream_synthesize(
        request.text,
        request.ref_audio,
        request.ref_text,
        model_global,
        vocoder_global,
        nfe_step=request.nfe_step,
        device=DEVICE,
    ):
        chunks.append(chunk)

    full_audio = b"".join(chunks)
    media_type = "audio/ogg" if request.output_format == "opus" else "audio/wav"
    return StreamingResponse(io.BytesIO(full_audio), media_type=media_type)


@app.get("/health")
async def health():
    """Health check endpoint."""
    return {"status": "ok", "model_loaded": model_global is not None}


@app.get("/info")
async def info():
    """Server info."""
    return {
        "model": "Habibi-TTS ALG (Specialized)",
        "base_model": "F5-TTS v1 Base",
        "device": DEVICE,
        "optimizations": ["BF16", "torch.compile", "EPSS"],
        "default_nfe": 7,
        "sample_rate": TARGET_SAMPLE_RATE,
    }


# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------


def main():
    parser = argparse.ArgumentParser(description="Habibi-TTS ALG Streaming Server")
    parser.add_argument("--host", default="0.0.0.0")
    parser.add_argument("--port", type=int, default=8000)
    parser.add_argument("--workers", type=int, default=1)
    args = parser.parse_args()

    import uvicorn
    uvicorn.run(app, host=args.host, port=args.port, workers=args.workers)


if __name__ == "__main__":
    main()