speech

Wav2Vec2-Large-Tedlium

The Wav2Vec2 large model fine-tuned on the TEDLIUM corpus.

The model is initialised with Facebook's Wav2Vec2 large LV-60k checkpoint pre-trained on 60,000h of audiobooks from the LibriVox project. It is fine-tuned on 452h of TED talks from the TEDLIUM corpus (Release 3). When using the model, make sure that your speech input is sampled at 16Khz.

The model achieves a word error rate (WER) of 8.4% on the dev set and 8.2% on the test set. Training logs document the training and evaluation progress over 50k steps of fine-tuning.

See this notebook for more information on how this model was fine-tuned.

Usage

To transcribe audio files the model can be used as a standalone acoustic model as follows:

 from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
 from datasets import load_dataset
 import torch
 
 # load model and processor
 processor = Wav2Vec2Processor.from_pretrained("sanchit-gandhi/wav2vec2-large-tedlium")
 model = Wav2Vec2ForCTC.from_pretrained("sanchit-gandhi/wav2vec2-large-tedlium")
     
 # load dummy dataset
 ds = load_dataset("sanchit-gandhi/tedlium_dummy", split="validation")
 
 # process audio inputs
 input_values = processor(ds[0]["audio"]["array"], return_tensors="pt", padding="longest").input_values  # Batch size 1
 
 # retrieve logits
 logits = model(input_values).logits
 
 # take argmax and decode
 predicted_ids = torch.argmax(logits, dim=-1)
 transcription = processor.batch_decode(predicted_ids)
 print("Target: ", ds["text"][0])
 print("Transcription: ", transcription[0])

Evaluation

This code snippet shows how to evaluate Wav2Vec2-Large-Tedlium on the TEDLIUM test data.

from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import torch
from jiwer import wer

tedlium_eval = load_dataset("LIUM/tedlium", "release3", split="test")
model = Wav2Vec2ForCTC.from_pretrained("sanchit-gandhi/wav2vec2-large-tedlium").to("cuda")
processor = Wav2Vec2Processor.from_pretrained("sanchit-gandhi/wav2vec2-large-tedlium")
def map_to_pred(batch):
    input_values = processor(batch["audio"]["array"], return_tensors="pt", padding="longest").input_values
    with torch.no_grad():
        logits = model(input_values.to("cuda")).logits
    predicted_ids = torch.argmax(logits, dim=-1)
    transcription = processor.batch_decode(predicted_ids)
    batch["transcription"] = transcription
    return batch
result = tedlium_eval.map(map_to_pred, batched=True, batch_size=1, remove_columns=["speech"])
print("WER:", wer(result["text"], result["transcription"]))