facebook/unit_hifigan_HK_layer12.km2500_frame_TAT-TTS

unit_hifigan_HK_layer12.km2500_frame_TAT-TTS

Hokkien unit HiFiGAN based vocoder from fairseq:

• Trained with TAT-TTS data with 4 speakers in Taiwanese Hokkien accent. See here
  for more training details.

Usage

import json<br /> import os<br /> from pathlib import Path<br /> import IPython.display as ipd<br /> from fairseq import hub_utils<br /> from fairseq.checkpoint_utils import load_model_ensemble_and_task_from_hf_hub<br /> from fairseq.models.speech_to_text.hub_interface import S2THubInterface<br /> from fairseq.models.text_to_speech import CodeHiFiGANVocoder<br /> from fairseq.models.text_to_speech.hub_interface import VocoderHubInterface<br /> from huggingface_hub import snapshot_download<br /> import torchaudio<br /> cache_dir = os.getenv("HUGGINGFACE_HUB_CACHE")<br /> # speech synthesis<br /> library_name = "fairseq"<br /> cache_dir = (<br /> cache_dir or (Path.home() / ".cache" / library_name).as_posix()<br /> )<br /> cache_dir = snapshot_download(<br /> f"facebook/unit_hifigan_HK_layer12.km2500_frame_TAT-TTS", cache_dir=cache_dir, library_name=library_name<br /> )<br /> x = hub_utils.from_pretrained(<br /> cache_dir,<br /> "model.pt",<br /> ".",<br /> archive_map=CodeHiFiGANVocoder.hub_models(),<br /> config_yaml="config.json",<br /> fp16=False,<br /> is_vocoder=True,<br /> )<br /> with open(f"{x['args']['data']}/config.json") as f:<br /> vocoder_cfg = json.load(f)<br /> assert (<br /> len(x["args"]["model_path"]) == 1<br /> ), "Too many vocoder models in the input"<br /> vocoder = CodeHiFiGANVocoder(x["args"]["model_path"][0], vocoder_cfg)<br /> tts_model = VocoderHubInterface(vocoder_cfg, vocoder)<br /> tts_sample = tts_model.get_model_input(unit)<br /> wav, sr = tts_model.get_prediction(tts_sample)<br /> ipd.Audio(wav, rate=sr)<br />