rinna / nue-asr

rinna/nue-asr

Overview

[Paper] [GitHub]

We propose a novel end-to-end speech recognition model, Nue ASR , which integrates pre-trained speech and language models.

The name Nue comes from the Japanese word ( 鵺/ぬえ/Nue ), one of the Japanese legendary creatures ( 妖怪/ようかい/Yōkai ).

This model provides end-to-end Japanese speech recognition with recognition accuracy comparable to the recent ASR models. You can recognize speech faster than real time by using a GPU.

Benchmark scores, including our models, can be found at https://rinnakk.github.io/research/benchmarks/asr/

• Model architecture

  This model consists of three main components: HuBERT audio encoder, bridge network, and GPT-NeoX decoder. The weights of HuBERT and GPT-NeoX were initialized with the pre-trained weights of HuBERT and GPT-NeoX, respectively.

  • japanese-hubert-base
  • japanese-gpt-neox-3.6b

• Training

  The model was trained on approximately 19,000 hours of following Japanese speech corpus ReazonSpeech v1. Note that speech samples longer than 16 seconds were excluded before training.

  • ReazonSpeech

• Contributors

  • Yukiya Hono
  • Koh Mitsuda
  • Tianyu Zhao
  • Kentaro Mitsui
  • Toshiaki Wakatsuki
  • Kei Sawada

How to use the model

We tested our code using Python 3.8.10 and 3.10.12 with PyTorch 2.1.1 and Transformers 4.35.2. This codebase is expected to be compatible with Python 3.8 or later and recent PyTorch versions. The version of Transformers should be 4.33.0 or higher.

First, install the code for inference of this model.

pip install git+https://github.com/rinnakk/nue-asr.git

Command-line interface and python interface are available.

Command-line usage

The following command transcribes the audio file using the command line interface. Audio files will be automatically downsampled to 16kHz.

nue-asr audio1.wav

You can specify multiple audio files.

nue-asr audio1.wav audio2.flac audio3.mp3

We can use DeepSpeed-Inference to accelerate the inference speed of GPT-NeoX module. If you use DeepSpeed-Inference, you need to install DeepSpeed.

pip install deepspeed

Then, you can use DeepSpeed-Inference as follows:

nue-asr --use-deepspeed audio1.wav

Run nue-asr --help for more information.

Python usage

The example of Python interface is as follows:

import nue_asr

model = nue_asr.load_model("rinna/nue-asr")
tokenizer = nue_asr.load_tokenizer("rinna/nue-asr")

result = nue_asr.transcribe(model, tokenizer, "path_to_audio.wav")
print(result.text)

nue_asr.transcribe function can accept audio data as either a numpy.array or a torch.Tensor , in addition to audio file paths.

Acceleration of inference speed using DeepSpeed-Inference is also available within the Python interface.

import nue_asr

model = nue_asr.load_model("rinna/nue-asr", use_deepspeed=True)
tokenizer = nue_asr.load_tokenizer("rinna/nue-asr")

result = nue_asr.transcribe(model, tokenizer, "path_to_audio.wav")
print(result.text)

Tokenization

The model uses the same sentencepiece-based tokenizer as japanese-gpt-neox-3.6b .

How to cite

@inproceedings{hono2024integrating,
    title = {Integrating Pre-Trained Speech and Language Models for End-to-End Speech Recognition},
    author = {Hono, Yukiya and Mitsuda, Koh and Zhao, Tianyu and Mitsui, Kentaro and Wakatsuki, Toshiaki and Sawada, Kei},
    booktitle = {Findings of the Association for Computational Linguistics: ACL 2024},
    year = {2024}
}

@misc{rinna-nue-asr,
    title = {rinna/nue-asr},
    author = {Hono, Yukiya and Mitsuda, Koh and Zhao, Tianyu and Mitsui, Kentaro and Wakatsuki, Toshiaki and Sawada, Kei},
    url = {https://huggingface.co/rinna/nue-asr}
}

References

@inproceedings{sawada2024release,
    title = {Release of Pre-Trained Models for the {J}apanese Language},
    author = {Sawada, Kei and Zhao, Tianyu and Shing, Makoto and Mitsui, Kentaro and Kaga, Akio and Hono, Yukiya and Wakatsuki, Toshiaki and Mitsuda, Koh},
    booktitle = {Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
    month = {5},
    year = {2024},
    pages = {13898--13905},
    url = {https://aclanthology.org/2024.lrec-main.1213},
    note = {\url{https://arxiv.org/abs/2404.01657}}
}

@article{hsu2021hubert,
    title = {{HuBERT}: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units},
    author = {Hsu, Wei-Ning and Bolte, Benjamin and Tsai, Yao-Hung Hubert and Lakhotia, Kushal and Salakhutdinov, Ruslan and Mohamed, Abdelrahman},
    journal = {IEEE/ACM Transactions on Audio, Speech, and Language Processing},
    month = {10},
    year = {2021},
    volume = {29},
    pages = {3451-3460},
    doi = {10.1109/TASLP.2021.3122291}
}

@software{andoniangpt2021gpt,
    title = {{GPT}-{N}eo{X}: Large Scale Autoregressive Language Modeling in {P}y{T}orch},
    author = {Andonian, Alex and Anthony, Quentin and Biderman, Stella and Black, Sid and Gali, Preetham and Gao, Leo and Hallahan, Eric and Levy-Kramer, Josh and Leahy, Connor and Nestler, Lucas and Parker, Kip and Pieler, Michael and Purohit, Shivanshu and Songz, Tri and Phil, Wang and Weinbach, Samuel},
    month = {8},
    year = {2021},
    version = {0.0.1},
    doi = {10.5281/zenodo.5879544},
    url = {https://www.github.com/eleutherai/gpt-neox}
}

@inproceedings{aminabadi2022deepspeed,
    title = {{DeepSpeed-Inference}: enabling efficient inference of transformer models at unprecedented scale},
    author = {Aminabadi, Reza Yazdani and Rajbhandari, Samyam and Awan, Ammar Ahmad and Li, Cheng and Li, Du and Zheng, Elton and Ruwase, Olatunji and Smith, Shaden and Zhang, Minjia and Rasley, Jeff and others},
    booktitle = {SC22: International Conference for High Performance Computing, Networking, Storage and Analysis},
    year = {2022},
    pages = {1--15},
    doi = {10.1109/SC41404.2022.00051}
}

License

The Apache 2.0 license