metadata
thumbnail: https://github.com/rinnakk/japanese-pretrained-models/blob/master/rinna.png
language: ja
license: apache-2.0
datasets: reazon-research/reazonspeech
inference: false
tags:
- data2vec
- speech
rinna/japanese-data2vec-audio-base
Overview
This is a Japanese data2vec Audio Base model trained by rinna Co., Ltd.
Model summary
The model architecture is the same as the original data2vec Audio Base model, which contains 12 transformer layers with 12 attention heads. The model was trained using code from the official repository, and the detailed training configuration can be found in the same repository and the original paper.
Training
The model was trained on approximately 19,000 hours of following Japanese speech corpus ReazonSpeech v1.
Contributors
How to use the model
import soundfile as sf
from transformers import AutoFeatureExtractor, AutoModel
model_name = "rinna/japanese-data2vec-audio-base"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)
model.eval()
raw_speech_16kHz, sr = sf.read(audio_file)
inputs = feature_extractor(
raw_speech_16kHz,
return_tensors="pt",
sampling_rate=sr,
)
outputs = model(**inputs)
print(f"Input: {inputs.input_values.size()}") # [1, #samples]
print(f"Output: {outputs.last_hidden_state.size()}") # [1, #frames, 768]
A fairseq checkpoint file can also be available here.
How to cite
@misc{rinna-japanese-data2vec-audio-base,
title={rinna/japanese-data2vec-audio-base},
author={Hono, Yukiya and Mitsui, Kentaro and Sawada, Kei},
url={https://huggingface.co/rinna/japanese-data2vec-audio-base}
}
Citations
@inproceedings{baevski2022data2vec,
title={Data2vec: A general framework for self-supervised learning in speech, vision and language},
author={Baevski, Alexei and Hsu, Wei-Ning and Xu, Qiantong and Babu, Arun and Gu, Jiatao and Auli, Michael},
booktitle={International Conference on Machine Learning},
pages={1298--1312},
year={2022},
organization={PMLR},
doi={10.48550/arXiv.2202.03555}
}