wav2vec2-librispeech-en-finetuned

wav2vec2-librispeech-en-finetuned

Fine-tuned facebook/wav2vec2-large-xlsr-53 for English ASR using LibriSpeech (openslr/librispeech_asr, clean, train.360).

• Task: Automatic Speech Recognition (CTC)
• Language: English
• License: CC BY 4.0
• Base model: facebook/wav2vec2-large-xlsr-53
• Dataset: openslr/librispeech_asr
• Validation WER (best): 0.0691
• This model is uncased: outputs are lowercased and generally do not restore punctuation or capitalization.

Usage

import torch
import librosa
from transformers import AutoProcessor, AutoModelForCTC

repo_id = "takehika/wav2vec2-librispeech-en-finetuned"
processor = AutoProcessor.from_pretrained(repo_id)
model = AutoModelForCTC.from_pretrained(repo_id)

speech, _ = librosa.load("sample.wav", sr=16_000)
inputs = processor(speech, sampling_rate=16_000, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

pred_ids = torch.argmax(logits, dim=-1)
text = processor.batch_decode(pred_ids)[0]
print(text)

Input vs Output Example

• Spoken content: HELLO, WORLD. I'M JOHN.
• Model output style: hello world im john

Data

• Dataset: openslr/librispeech_asr
• Config: clean
• Train: train.360 (104,014 samples)
• Validation: validation (2,703 samples)
• Test: test (2,620 samples)

Training

• Batch size: 8
• Gradient accumulation: 2 (effective 16)
• Learning rate: 2e-5
• Scheduler: constant_with_warmup
• Warmup steps: 500
• Eval/save steps: 500

Evaluation

• Validation WER: 0.0691

Intended Use & Limitations

• Intended use: English ASR for 16 kHz speech close to read/clean speech conditions.
• Domain shift: performance may degrade on noisy audio, strong accents, telephone speech, overlapping speakers, or code-switching.
• Output style: due to CTC + lowercase transcript setup, punctuation and casing quality can be limited.
• Reliability: transcription errors are possible; human review is recommended for high-stakes use.

Attribution & Licenses

• License: CC BY 4.0
• Base model facebook/wav2vec2-large-xlsr-53: Apache-2.0
• Dataset openslr/librispeech_asr: CC BY 4.0

References:

• Base model card: https://huggingface.co/facebook/wav2vec2-large-xlsr-53
• Dataset card: https://huggingface.co/datasets/openslr/librispeech_asr
• LibriSpeech source (OpenSLR SLR12): https://openslr.org/12/

Dataset Citation

@inproceedings{panayotov2015librispeech,
  title={Librispeech: an ASR corpus based on public domain audio books},
  author={Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev},
  booktitle={Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on},
  pages={5206--5210},
  year={2015},
  organization={IEEE}
}