Skip to content

SOTA Open Source TTS

10.1 KB
Raw Blame
import argparse
import base64
import time
import wave

import ormsgpack
import pyaudio
import requests
from pydub import AudioSegment
from pydub.playback import play

from fish_speech.utils.file import audio_to_bytes, read_ref_text
from fish_speech.utils.schema import ServeReferenceAudio, ServeTTSRequest


def parse_args():
    parser = argparse.ArgumentParser(
        description="Send text to a Fish Speech TTS server and receive synthesized audio.",
        epilog=(
            "Model selection note:\n"
            "  The base TTS model is selected by the server you call. For example, if the\n"
            "  server was started with checkpoints/s2-pro, this client will use S2-Pro\n"
            "  automatically. There is no separate per-request --model flag.\n\n"
            "Examples:\n"
            '  python tools/api_client.py -u http://127.0.0.1:8080/v1/tts -t "Hello from Fish Speech"\n'
            '  python tools/api_client.py -u http://127.0.0.1:8080/v1/tts -t "Hello" --reference_id my-speaker'
        ),
        formatter_class=argparse.RawTextHelpFormatter,
    )

    parser.add_argument(
        "--url",
        "-u",
        type=str,
        default="http://127.0.0.1:8080/v1/tts",
        help="URL of the TTS server. The server decides which base model is loaded.",
    )
    parser.add_argument(
        "--text", "-t", type=str, required=True, help="Text to be synthesized"
    )
    parser.add_argument(
        "--reference_id",
        "-id",
        type=str,
        default=None,
        help="ID of the reference voice to use for synthesis\n(Local: name of folder containing audios and files)",
    )
    parser.add_argument(
        "--reference_audio",
        "-ra",
        type=str,
        nargs="+",
        default=None,
        help="Path to the audio file",
    )
    parser.add_argument(
        "--reference_text",
        "-rt",
        type=str,
        nargs="+",
        default=None,
        help="Reference text for voice synthesis",
    )
    parser.add_argument(
        "--reference_speaker",
        "-rs",
        type=int,
        nargs="+",
        default=None,
        help="Speaker ID (0-15) of each reference audio, written as <|speaker:N|> "
        "before reference texts that do not start with a speaker tag",
    )
    parser.add_argument(
        "--output",
        "-o",
        type=str,
        default="generated_audio",
        help="Output audio file name",
    )
    parser.add_argument(
        "--play",
        action=argparse.BooleanOptionalAction,
        default=True,
        help="Whether to play audio after receiving data",
    )
    parser.add_argument(
        "--format", type=str, choices=["wav", "pcm", "mp3", "opus"], default="wav"
    )
    parser.add_argument(
        "--latency",
        type=str,
        default="normal",
        choices=["normal", "balanced"],
        help="Used in api.fish.audio/v1/tts",
    )
    # Sampling: unset means the server uses the loaded model family's defaults.
    parser.add_argument(
        "--max_new_tokens",
        type=int,
        default=None,
        help="Maximum new tokens to generate per text chunk. Unset uses the "
        "server's default, 1024; 0 means the model's own cap (S2-Pro: the context "
        "length; Drama 3: 600 per chunk, 2048 with --single_chunk).",
    )
    parser.add_argument(
        "--single_chunk",
        action="store_true",
        help="Drama 3 only: generate the whole text in one pass instead of chunks",
    )
    parser.add_argument(
        "--chunk_length",
        type=int,
        default=200,
        help="Chunk length for synthesis (S2-Pro only; Drama 3 uses ~300-byte chunks)",
    )
    parser.add_argument(
        "--top_p", type=float, default=None, help="Top-p sampling for synthesis"
    )
    parser.add_argument(
        "--repetition_penalty",
        type=float,
        default=None,
        help="Accepted for compatibility; S2-Pro and Drama 3 ignore it and use "
        "repetition-aware sampling",
    )
    parser.add_argument(
        "--temperature", type=float, default=None, help="Temperature for sampling"
    )

    # parser.add_argument("--streaming", type=bool, default=False, help="Enable streaming response")
    parser.add_argument(
        "--streaming", action="store_true", help="Enable streaming response"
    )
    parser.add_argument(
        "--channels", type=int, default=1, help="Number of audio channels"
    )
    parser.add_argument("--rate", type=int, default=44100, help="Sample rate for audio")
    parser.add_argument(
        "--use_memory_cache",
        type=str,
        default="off",
        choices=["on", "off"],
        help="Cache encoded references codes in memory.\n",
    )
    parser.add_argument(
        "--seed",
        type=int,
        default=None,
        help="`None` means randomized inference, otherwise deterministic.\nIt can't be used for fixing a timbre.",
    )
    parser.add_argument(
        "--api_key",
        type=str,
        default="YOUR_API_KEY",
        help="API key for authentication",
    )

    return parser.parse_args()


if __name__ == "__main__":
    args = parse_args()

    idstr: str | None = args.reference_id
    # priority: ref_id > [{text, audio},...]
    if idstr is None:
        ref_audios = args.reference_audio or []
        ref_texts = args.reference_text or []

        if len(ref_audios) != len(ref_texts):
            raise SystemExit(
                "--reference_audio and --reference_text must be given the same number "
                f"of values, got {len(ref_audios)} audio(s) and {len(ref_texts)} text(s)."
            )

        byte_audios = []
        for ref_audio in ref_audios:
            audio_bytes = audio_to_bytes(ref_audio)
            if audio_bytes is None:
                raise SystemExit(f"Reference audio file not found: {ref_audio}")
            byte_audios.append(audio_bytes)

        ref_texts = [read_ref_text(ref_text) for ref_text in ref_texts]
        ref_speakers = args.reference_speaker or [None] * len(ref_audios)
        if len(ref_speakers) != len(ref_audios):
            raise SystemExit(
                "--reference_speaker must be given once per --reference_audio, got "
                f"{len(ref_speakers)} speaker(s) and {len(ref_audios)} audio(s)."
            )
    else:
        byte_audios = []
        ref_texts = []
        ref_speakers = []
        pass  # in api.py

    data = {
        "text": args.text,
        "references": [
            ServeReferenceAudio(audio=ref_audio, text=ref_text, speaker=speaker)
            for ref_text, ref_audio, speaker in zip(
                ref_texts, byte_audios, ref_speakers
            )
        ],
        "reference_id": idstr,
        "format": args.format,
        "latency": args.latency,
        "max_new_tokens": args.max_new_tokens,
        "chunk_length": args.chunk_length,
        "single_chunk": args.single_chunk,
        "top_p": args.top_p,
        "repetition_penalty": args.repetition_penalty,
        "temperature": args.temperature,
        "streaming": args.streaming,
        "use_memory_cache": args.use_memory_cache,
        "seed": args.seed,
    }

    pydantic_data = ServeTTSRequest(**data)

    print("Sending request")
    start_time = time.time()
    response = requests.post(
        args.url,
        params={"format": "msgpack"},
        # Unset (None) fields are left out so servers that predate optional sampling
        # fields fall back to their own defaults instead of rejecting null.
        data=ormsgpack.packb(pydantic_data.model_dump(exclude_none=True)),
        stream=args.streaming,
        headers={
            "authorization": f"Bearer {args.api_key}",
            "content-type": "application/msgpack",
        },
    )
    end_time = time.time()
    print(f"Request took {end_time - start_time} seconds")

    if response.status_code == 200:
        if args.streaming:
            p = pyaudio.PyAudio()
            audio_format = pyaudio.paInt16  # Assuming 16-bit PCM format
            stream = p.open(
                format=audio_format, channels=args.channels, rate=args.rate, output=True
            )

            wf = wave.open(f"{args.output}.wav", "wb")
            wf.setnchannels(args.channels)
            wf.setsampwidth(p.get_sample_size(audio_format))
            wf.setframerate(args.rate)

            stream_stopped_flag = False
            # The server opens the stream with a 44-byte WAV header (older ones
            # did not); only the samples after it are played and saved.
            header = b""

            try:
                for chunk in response.iter_content(chunk_size=1024):
                    if header is not None:
                        header += chunk
                        if len(header) < 44:
                            continue
                        chunk = header[44:] if header.startswith(b"RIFF") else header
                        header = None
                    if chunk:
                        stream.write(chunk)
                        wf.writeframesraw(chunk)
                    else:
                        if not stream_stopped_flag:
                            stream.stop_stream()
                            stream_stopped_flag = True
            finally:
                stream.close()
                p.terminate()
                wf.close()
        else:
            audio_content = response.content
            audio_path = f"{args.output}.{args.format}"
            with open(audio_path, "wb") as audio_file:
                audio_file.write(audio_content)

            if args.format == "pcm":  # headerless 16-bit mono
                audio = AudioSegment.from_file(
                    audio_path,
                    format="raw",
                    sample_width=2,
                    frame_rate=args.rate,
                    channels=1,
                )
            else:  # opus comes in an Ogg container
                audio = AudioSegment.from_file(
                    audio_path, format="ogg" if args.format == "opus" else args.format
                )
            if args.play:
                play(audio)
            print(f"Audio has been saved to '{audio_path}'.")
    else:
        print(f"Request failed with status code {response.status_code}")
        print(response.json())