Wav2Vec2-Base-960h

Wav2Vec2-Base-960h

Turns English speech audio sampled at 16 kHz into text. A base wav2vec 2.0 model fine-tuned on 960 hours of Librispeech, for developers and researchers who need a transcription baseline.

Made by facebook on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Wav2Vec2-Base-960h is a speech-to-text model for English. It reads an audio clip and writes out the words. Facebook pretrained the base model on speech audio and then fine-tuned it on 960 hours of Librispeech.

Who it suits: It suits developers and researchers who need English transcription or a baseline to compare against. People have published 181 fine-tuned versions of it.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Software | Developers | Add English transcription to an app | A script that transcribes audio files
Research | Speech researchers | Use it as a baseline and measure word error rate | Scores on the clean and other LibriSpeech test sets
Education | Students of speech recognition | Study how a speech model turns audio into text | Run the card's sample on a small dataset

If your audio is not English, you can skip it.

Where to use it:

• Python on Windows, macOS or Linux
• The transformers library, with PyTorch
• A GPU, if you run the card's evaluation script, which sends the model to CUDA
• Any audio source, as long as it is sampled at 16 kHz

Title: Jobs it does
Use case | Who | What they get back
Transcribing audio files | Developers | Text decoded from the audio clip
Measuring accuracy | Researchers | A word error rate on LibriSpeech test data

How to set it up:

• Python on Windows, macOS or Linux
• The transformers, datasets and torch packages, which the card's code imports
• Speech input sampled at 16 kHz

1. Install the packages the code imports.
2. Optionally download the model files by hand.
3. Run the usage code, which loads the model by its name.

pip install -U huggingface_hub
hf download facebook/wav2vec2-base-960h

How to use it: Load the processor and the model, turn the audio array into input values, take the most likely token at each step, and decode the ids into text. The card shows this on a small sample dataset.

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

# load model and tokenizer
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")

# load dummy dataset and read soundfiles
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")

# tokenize
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)

Title: Pieces of the code
Tool | What it does
Wav2Vec2Processor | Turns the audio array into input values for the model
Wav2Vec2ForCTC | Runs the model and returns scores for each step
torch.argmax | Picks the most likely token at every step
batch_decode | Turns the chosen token ids into text

Example: You load the first clip of the card's small sample dataset and run the code above. The model reads the audio and gives back its transcription as a text string, which you can compare with the clip's reference text.

For newcomers and experts: If you are new, think of this as a program that listens to a recording and types what it hears. The first small thing to try is the sample code on the one clip from the card's dataset.

If you know the field, wav2vec 2.0 pretrains by masking parts of the audio features, then solving a contrastive task on quantised latent values. This checkpoint is used as a standalone acoustic model with CTC and no language model. The trade-off is accuracy on harder audio, with a word error rate of 3.4 on clean and 8.6 on other.

Audio sampled at 16 kHz
Processor makes input values
Wav2Vec2ForCTC returns logits
Argmax picks token ids
batch_decode writes the text

Good to know:

Title: Word error rate on LibriSpeech test data
clean: 3.4%
other: 8.6%

• Input: the audio must be sampled at 16 kHz.
• Language: the record lists English only.
• Size: 94.4M parameters, about 378 MB in F32.
• License: apache-2.0.
• Cost to run: no API key, account or paid service is named.
• Data it reaches: it downloads the model and a sample dataset from Hugging Face, and the card names no other access.
• Not for: speech in languages other than English.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 94.4M
• Weights: about 378 MB in F32; about 47 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,794,378
• Likes: 418
• License: apache-2.0
• Languages: en
• Quantized versions: 11
• Fine-tunes: 181
• Tags: transformers, pytorch, safetensors, wav2vec2, automatic-speech-recognition, audio, hf-asr-leaderboard, model-index

Links

• Model: https://huggingface.co/facebook/wav2vec2-base-960h
• Files: https://huggingface.co/facebook/wav2vec2-base-960h/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download facebook/wav2vec2-base-960h

Then follow the usage in its model card. Nothing in this kit runs by itself.