🚀 Wav2Vec2-Large-960h-Lv60 + Self-Training
A large model pretrained and fine-tuned on 960 hours of Libri-Light and Librispeech for automatic speech recognition.
This is a large model that has been pretrained and fine-tuned on 960 hours of Libri-Light and Librispeech speech audio sampled at 16kHz. It was trained with the Self-Training objective. When using this model, ensure that your speech input is also sampled at 16kHz.
Model Information
Property |
Details |
Model Type |
wav2vec2-large-960h-lv60 |
Training Data |
Libri-Light and Librispeech (960 hours, 16kHz sampled speech audio) |
License |
apache-2.0 |
Tags |
speech, audio, automatic-speech-recognition, hf-asr-leaderboard |
Results
Task |
Dataset |
Test WER |
Automatic Speech Recognition |
LibriSpeech (clean) |
1.9 |
Automatic Speech Recognition |
LibriSpeech (other) |
3.9 |
Paper and Authors
- Paper
- Authors: Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli
Abstract
The paper shows for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
🚀 Quick Start
To get started with this model, follow the steps below.
💻 Usage Examples
Basic Usage
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torch
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-large-960h-lv60-self")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h-lv60-self")
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
input_values = processor(ds[0]["audio"]["array"], return_tensors="pt", padding="longest").input_values
logits = model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)
Evaluation
This code snippet shows how to evaluate facebook/wav2vec2-large-960h-lv60-self on LibriSpeech's "clean" and "other" test data.
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import torch
from jiwer import wer
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h-lv60-self").to("cuda")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-large-960h-lv60-self")
def map_to_pred(batch):
inputs = processor(batch["audio"]["array"], return_tensors="pt", padding="longest")
input_values = inputs.input_values.to("cuda")
attention_mask = inputs.attention_mask.to("cuda")
with torch.no_grad():
logits = model(input_values, attention_mask=attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)
batch["transcription"] = transcription
return batch
result = librispeech_eval.map(map_to_pred, remove_columns=["audio"])
print("WER:", wer(result["text"], result["transcription"]))
Result (WER):