mah92/vocos-matcha-16khz

Vocos Matcha 16 kHz (80 mels) — sherpa-onnx vocoder

Vocos vocoder by Ali Mahmoudi (@mah92), trained to pair with Matcha-TTS 16 kHz models, exported for sherpa-onnx.

[image]

Matcha compatibility

The well-known vocos-16khz-univ.onnx is not compatible with Matcha models — it was trained on a different mel style, and produces near-silent audio with Matcha mels.

This model was trained on the same mel style Matcha uses (slaney norm, slaney mel scale, fmin 0, fmax 8000, 80 bins, 16 kHz), so it works with Matcha mels directly — no tricks needed.

Training

• —Repo: bsc-vocos (langtech-bsc/vocos@matcha fork, with the matcha-mel patch)
• —Data: LibriSpeech 16 kHz
• —Best checkpoint: step 964,650 (epoch 58), val_loss 3.9397, UTMOS ≈ 3.37, PESQ ≈ 2.6

Usage with sherpa-onnx

import sherpa_onnx

tts = sherpa_onnx.OfflineTts(sherpa_onnx.OfflineTtsConfig(
    model=sherpa_onnx.OfflineTtsModelConfig(
        matcha=sherpa_onnx.OfflineTtsMatchaModelConfig(
            acoustic_model="matcha-fa_en-zahra-16000-5.onnx",
            vocoder="vocos-matcha-16khz.onnx",
            tokens="tokens.txt",
            data_dir="espeak-ng-data",
            noise_scale=1.0,
            length_scale=1.0,
        ),
        num_threads=4,
    )
))
audio = tts.generate("سلام، این یک آزمایش است.", sid=0, speed=1.0)

Files

• —vocos-matcha-16khz.onnx — vocoder ONNX
• —vocos-matcha-16KHz.yaml — training config
• —vocos16_loss_curves.png — training curves

Pairing rule

Acoustic model must be Matcha-TTS, 16 kHz, 80 mels (slaney) — e.g. mah92/Zahra-FA_EN-16KHz-Matcha-TTS-Model

Author

Ali Mahmoudi (@mah92)