app.py

from functools import lru_cache

import sherpa_onnx
from huggingface_hub import hf_hub_download


def get_file(
    repo_id: str,
    filename: str,
    subfolder: str = ".",
) -> str:
    model_filename = hf_hub_download(
        repo_id=repo_id,
        filename=filename,
        subfolder=subfolder,
    )
    return model_filename


@lru_cache(maxsize=10)
def _get_vits_vctk(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
    assert repo_id == "csukuangfj/vits-vctk"

    model = get_file(
        repo_id=repo_id,
        filename="vits-vctk.onnx",
        subfolder=".",
    )

    lexicon = get_file(
        repo_id=repo_id,
        filename="lexicon.txt",
        subfolder=".",
    )

    tokens = get_file(
        repo_id=repo_id,
        filename="tokens.txt",
        subfolder=".",
    )

    tts_config = sherpa_onnx.OfflineTtsConfig(
        model=sherpa_onnx.OfflineTtsModelConfig(
            vits=sherpa_onnx.OfflineTtsVitsModelConfig(
                model=model,
                lexicon=lexicon,
                tokens=tokens,
                length_scale=1.0 / speed,
            ),
            provider="cpu",
            debug=True,
            num_threads=2,
        )
    )
    tts = sherpa_onnx.OfflineTts(tts_config)

    return tts


@lru_cache(maxsize=10)
def _get_vits_ljs(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
    assert repo_id == "csukuangfj/vits-ljs"

    model = get_file(
        repo_id=repo_id,
        filename="vits-ljs.onnx",
        subfolder=".",
    )

    lexicon = get_file(
        repo_id=repo_id,
        filename="lexicon.txt",
        subfolder=".",
    )

    tokens = get_file(
        repo_id=repo_id,
        filename="tokens.txt",
        subfolder=".",
    )

    tts_config = sherpa_onnx.OfflineTtsConfig(
        model=sherpa_onnx.OfflineTtsModelConfig(
            vits=sherpa_onnx.OfflineTtsVitsModelConfig(
                model=model,
                lexicon=lexicon,
                tokens=tokens,
                length_scale=1.0 / speed,
            ),
            provider="cpu",
            debug=True,
            num_threads=2,
        )
    )
    tts = sherpa_onnx.OfflineTts(tts_config)

    return tts


@lru_cache(maxsize=10)
def _get_vits_piper(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
    data_dir = "/tmp/espeak-ng-data"
    if "coqui" in repo_id or "vits-mms" in repo_id:
        name = "model"
    elif "piper" in repo_id:
        n = len("vits-piper-")
        name = repo_id.split("/")[1][n:]
    elif "mimic3" in repo_id:
        n = len("vits-mimic3-")
        name = repo_id.split("/")[1][n:]
    else:
        raise ValueError(f"Unsupported {repo_id}")

    if "vits-coqui-uk-mai" in repo_id or "vits-mms" in repo_id:
        data_dir = ""

    model = get_file(
        repo_id=repo_id,
        filename=f"{name}.onnx",
        subfolder=".",
    )

    tokens = get_file(
        repo_id=repo_id,
        filename="tokens.txt",
        subfolder=".",
    )

    tts_config = sherpa_onnx.OfflineTtsConfig(
        model=sherpa_onnx.OfflineTtsModelConfig(
            vits=sherpa_onnx.OfflineTtsVitsModelConfig(
                model=model,
                lexicon="",
                data_dir=data_dir,
                tokens=tokens,
                length_scale=1.0 / speed,
            ),
            provider="cpu",
            debug=True,
            num_threads=2,
        )
    )
    tts = sherpa_onnx.OfflineTts(tts_config)

    return tts


@lru_cache(maxsize=10)
def _get_vits_mms(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
    return _get_vits_piper(repo_id, speed)


@lru_cache(maxsize=10)
def _get_vits_zh_aishell3(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
    assert repo_id == "csukuangfj/vits-zh-aishell3"

    model = get_file(
        repo_id=repo_id,
        filename="vits-aishell3.onnx",
        subfolder=".",
    )

    lexicon = get_file(
        repo_id=repo_id,
        filename="lexicon.txt",
        subfolder=".",
    )

    tokens = get_file(
        repo_id=repo_id,
        filename="tokens.txt",
        subfolder=".",
    )

    rule_fst = get_file(
        repo_id=repo_id,
        filename="rule.fst",
        subfolder=".",
    )

    tts_config = sherpa_onnx.OfflineTtsConfig(
        model=sherpa_onnx.OfflineTtsModelConfig(
            vits=sherpa_onnx.OfflineTtsVitsModelConfig(
                model=model,
                lexicon=lexicon,
                tokens=tokens,
                length_scale=1.0 / speed,
            ),
            provider="cpu",
            debug=True,
            num_threads=2,
        ),
        rule_fsts=rule_fst,
    )
    tts = sherpa_onnx.OfflineTts(tts_config)

    return tts


@lru_cache(maxsize=10)
def _get_vits_hf(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
    if "fanchen" in repo_id or "vits-cantonese-hf-xiaomaiiwn" in repo_id:
        model = repo_id.split("/")[-1]
    else:
        model = repo_id.split("-")[-1]

    model = get_file(
        repo_id=repo_id,
        filename=f"{model}.onnx",
        subfolder=".",
    )

    lexicon = get_file(
        repo_id=repo_id,
        filename="lexicon.txt",
        subfolder=".",
    )

    tokens = get_file(
        repo_id=repo_id,
        filename="tokens.txt",
        subfolder=".",
    )

    rule_fst = get_file(
        repo_id=repo_id,
        filename="rule.fst",
        subfolder=".",
    )

    tts_config = sherpa_onnx.OfflineTtsConfig(
        model=sherpa_onnx.OfflineTtsModelConfig(
            vits=sherpa_onnx.OfflineTtsVitsModelConfig(
                model=model,
                lexicon=lexicon,
                tokens=tokens,
                length_scale=1.0 / speed,
            ),
            provider="cpu",
            debug=True,
            num_threads=2,
        ),
        rule_fsts=rule_fst,
    )
    tts = sherpa_onnx.OfflineTts(tts_config)

    return tts


@lru_cache(maxsize=10)
def get_pretrained_model(repo_id: str, speed: float) -> sherpa_onnx.OfflineTts:
    if repo_id in english_models:
        return english_models[repo_id](repo_id, speed)
    elif repo_id in arabic_models:
        return arabic_models[repo_id](repo_id, speed)
    elif repo_id in turkish_models:
        return turkish_models[repo_id](repo_id, speed)
    elif repo_id in persian_models:
        return persian_models[repo_id](repo_id, speed)
    elif repo_id in hindi_models:
        return hindi_models[repo_id](repo_id, speed)
    elif repo_id in gujarati_models:
        return gujarati_models[repo_id](repo_id, speed)
    else:
        raise ValueError(f"Unsupported repo_id: {repo_id}")


english_models = {
    "csukuangfj/vits-piper-en_US-glados": _get_vits_piper,
    # coqui-ai
    "csukuangfj/vits-coqui-en-ljspeech": _get_vits_piper,
    "csukuangfj/vits-coqui-en-ljspeech-neon": _get_vits_piper,
    "csukuangfj/vits-coqui-en-vctk": _get_vits_piper,
    # piper, US
    "csukuangfj/vits-piper-en_GB-sweetbbak-amy": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-amy-low": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-amy-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-arctic-medium": _get_vits_piper,  # 18 speakers
    "csukuangfj/vits-piper-en_US-danny-low": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-hfc_male-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-joe-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-kathleen-low": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-kusal-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-l2arctic-medium": _get_vits_piper,  # 24 speakers
    "csukuangfj/vits-piper-en_US-lessac-low": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-lessac-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-lessac-high": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-libritts-high": _get_vits_piper,  # 904 speakers
    "csukuangfj/vits-piper-en_US-libritts_r-medium": _get_vits_piper,  # 904 speakers
    "csukuangfj/vits-piper-en_US-ryan-low": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-ryan-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_US-ryan-high": _get_vits_piper,
    # piper, GB
    "csukuangfj/vits-piper-en_GB-alan-low": _get_vits_piper,
    "csukuangfj/vits-piper-en_GB-alan-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_GB-alba-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_GB-jenny_dioco-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_GB-northern_english_male-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_GB-semaine-medium": _get_vits_piper,
    "csukuangfj/vits-piper-en_GB-southern_english_female-low": _get_vits_piper,
    "csukuangfj/vits-piper-en_GB-vctk-medium": _get_vits_piper,
    #
    "csukuangfj/vits-vctk": _get_vits_vctk,  # 109 speakers
    "csukuangfj/vits-ljs": _get_vits_ljs,
}


arabic_models = {
    "csukuangfj/vits-piper-ar_JO-kareem-low": _get_vits_piper,
    "csukuangfj/vits-piper-ar_JO-kareem-medium": _get_vits_piper,
}

turkish_models = {
    "csukuangfj/vits-piper-tr_TR-dfki-medium": _get_vits_piper,
    "csukuangfj/vits-piper-tr_TR-fahrettin-medium": _get_vits_piper,
}


persian_models = {
    "csukuangfj/vits-piper-fa_IR-amir-medium": _get_vits_piper,
    "csukuangfj/vits-piper-fa_IR-gyro-medium": _get_vits_piper,
    "csukuangfj/vits-mimic3-fa-haaniye_low": _get_vits_piper,
}

gujarati_models = {
    "csukuangfj/vits-mimic3-gu_IN-cmu-indic_low": _get_vits_piper,
}
hindi_models = {
    "vosk-model-hi-0.22": _get_vits_piper,
}


language_to_models = {
    "English": list(english_models.keys()),
    "Arabic": list(arabic_models.keys()),
    "Hindi": list(hindi_models.keys()),
    "Gujarati": list(gujarati_models.keys()),
    "Persian": list(persian_models.keys()),
    "Turkish": list(turkish_models.keys()),
}