HiFi-GAN vocoder for Thorsten-Voice

`jjiang4/thorsten_hifigan_ft_ljspeech`

HiFi-GAN V1 vocoder for the German Thorsten-Voice Dataset 2022.10 speaker, fine-tuned for 40k steps from parallel_wavegan/ljspeech_hifigan.v1 with kan-bayashi/ParallelWaveGAN.

Intended for the ESPnet egs2/thorsten/tts1 recipe. It replaces Griffin-Lim and raises UTMOS from 1.42 to 3.04 on that recipe's test set.

Feature settings

Matches egs2/thorsten/tts1's feats_extract_conf exactly, so it consumes the recipe's mel-spectrograms unchanged:


sampling rate	22050
FFT size	1024
hop size	256
mel bins	80
fmin / fmax	80 / 7600

Files

checkpoint-40000steps.pkl — generator weights. Training state (discriminator and optimizers) is stripped; the output is bit-identical to the full checkpoint.
config.yml, stats.h5 — read by ESPnet when loading the vocoder. Keep all three files in the same directory.

Usage

cd espnet/egs2/thorsten/tts1
./run.sh --stage 8 --inference_model valid.loss.best.pth \
    --vocoder_file /path/to/checkpoint-40000steps.pkl

from espnet2.bin.tts_inference import Text2Speech
tts = Text2Speech.from_pretrained(
    "jjiang4/thorsten_tts_train_tacotron2_raw_phn_espeak_ng_german",
    vocoder_file="/path/to/checkpoint-40000steps.pkl",
)

Downloads last month: -; Downloads are not tracked for this model. How to track