|
--- |
|
language: |
|
- en |
|
tags: |
|
- audio |
|
- automatic-speech-recognition |
|
license: mit |
|
library_name: ctranslate2 |
|
--- |
|
|
|
# Distil-Whisper: distil-large-v3 for CTranslate2 |
|
|
|
This repository contains the model weights for [distil-large-v3](https://huggingface.co/distil-whisper/distil-large-v3) |
|
converted to [CTranslate2](https://github.com/OpenNMT/CTranslate2) format. CTranslate2 is a fast inference engine for |
|
Transformer models and is the supported backend for the [Faster-Whisper](https://github.com/systran/faster-whisper) package. |
|
|
|
Compared to previous Distil-Whisper releases, distil-large-v3 is specifically designed to be compatible |
|
with the OpenAI Whisper long-form transcription algorithm. In our benchmark over 4 out-of-distribution datasets, distil-large-v3 |
|
outperformed distil-large-v2 by 5% WER average. Thus, you can expect significant performance gains by switching to this |
|
latest checkpoint. |
|
|
|
## Usage |
|
|
|
To use the model in Faster-Whisper, first install the PyPi package according to the [official instructions](https://github.com/SYSTRAN/faster-whisper#installation). |
|
For this example, we'll also install 🤗 Datasets to load a toy audio dataset from the Hugging Face Hub: |
|
|
|
```bash |
|
pip install --upgrade pip |
|
pip install --upgrade git+https://github.com/SYSTRAN/faster-whisper datasets[audio] |
|
``` |
|
|
|
The following code snippet loads the distil-large-v3 model and runs inference on an example file from the LibriSpeech ASR |
|
dataset: |
|
|
|
```python |
|
import torch |
|
from faster_whisper import WhisperModel |
|
from datasets import load_dataset |
|
|
|
# define our torch configuration |
|
device = "cuda:0" if torch.cuda.is_available() else "cpu" |
|
compute_type = "float16" if torch.cuda.is_available() else "float32" |
|
|
|
# load model on GPU if available, else cpu |
|
model = WhisperModel("distil-large-v3", device=device, compute_type=compute_type) |
|
|
|
# load toy dataset for example |
|
dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation") |
|
sample = dataset[1]["audio"]["path"] |
|
|
|
segments, info = model.transcribe(sample, beam_size=1) |
|
|
|
for segment in segments: |
|
print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text)) |
|
``` |
|
|
|
To transcribe a local audio file, simply pass the path to the audio file as the `audio` argument to transcribe: |
|
|
|
```python |
|
segments, info = model.transcribe("audio.mp3", beam_size=1) |
|
``` |
|
|
|
## Model Details |
|
|
|
For more information about the distil-large-v3 model, refer to the original [model card](https://huggingface.co/distil-whisper/distil-large-v3). |
|
|
|
## License |
|
|
|
Distil-Whisper inherits the [MIT license](https://github.com/huggingface/distil-whisper/blob/main/LICENSE) from OpenAI's Whisper model. |
|
|
|
## Citation |
|
|
|
If you use this model, please consider citing the [Distil-Whisper paper](https://arxiv.org/abs/2311.00430): |
|
``` |
|
@misc{gandhi2023distilwhisper, |
|
title={Distil-Whisper: Robust Knowledge Distillation via Large-Scale Pseudo Labelling}, |
|
author={Sanchit Gandhi and Patrick von Platen and Alexander M. Rush}, |
|
year={2023}, |
|
eprint={2311.00430}, |
|
archivePrefix={arXiv}, |
|
primaryClass={cs.CL} |
|
} |
|
``` |
|
|
