Datasets documentation

Process audio data

You are viewing v2.0.0 version. A newer version v2.20.0 is available.
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Process audio data

🤗 Datasets supports an datasets.Audio feature, enabling users to load and process raw audio files for training. This guide will show you how to:


The datasets.Audio feature is an experimental feature and should be installed as an extra dependency in 🤗 Datasets. Install the datasets.Audio feature with pip:

>>> pip install datasets[audio]

Users should also install torchaudio and librosa, two common libraries used by 🤗 Datasets for handling audio data.

>>> pip install librosa
>>> pip install torchaudio

torchaudio’s sox_io backend supports decoding mp3 files. Unfortunately, the sox_io backend is only available on Linux/macOS, and is not supported by Windows.

Then you can load an audio dataset the same way you would load a text dataset. For example, load the Common Voice dataset with the Turkish configuration:

>>> from datasets import load_dataset, load_metric, Audio
>>> common_voice = load_dataset("common_voice", "tr", split="train")

Audio datasets

Audio datasets commonly have an audio and path or file column.

audio is the actual audio file that is loaded and resampled on-the-fly upon calling it.

>>> common_voice[0]["audio"]
{'array': array([ 0.0000000e+00,  0.0000000e+00,  0.0000000e+00, ...,
    -8.8930130e-05, -3.8027763e-05, -2.9146671e-05], dtype=float32),
'path': '/root/.cache/huggingface/datasets/downloads/extracted/05be0c29807a73c9b099873d2f5975dae6d05e9f7d577458a2466ecb9a2b0c6b/cv-corpus-6.1-2020-12-11/tr/clips/common_voice_tr_21921195.mp3',
'sampling_rate': 48000}

When you access an audio file, it is automatically decoded and resampled. Generally, you should query an audio file like: common_voice[0]["audio"]. If you query an audio file with common_voice["audio"][0] instead, all the audio files in your dataset will be decoded and resampled. This process can take a long time if you have a large dataset.

path or file is an absolute path to an audio file.

>>> common_voice[0]["path"]

The path is useful if you want to load your own audio dataset. In this case, provide a column of audio file paths to datasets.Dataset.cast_column():

>>> my_audio_dataset = my_audio_dataset.cast_column("paths_to_my_audio_files", Audio())


Some models expect the audio data to have a certain sampling rate due to how the model was pretrained. For example, the XLSR-Wav2Vec2 model expects the input to have a sampling rate of 16kHz, but an audio file from the Common Voice dataset has a sampling rate of 48kHz. You can use datasets.Dataset.cast_column() to downsample the sampling rate to 16kHz:

>>> common_voice = common_voice.cast_column("audio", Audio(sampling_rate=16_000))

The next time you load the audio file, the datasets.Audio feature will load and resample it to 16kHz:

>>> common_voice_train[0]["audio"]
{'array': array([ 0.0000000e+00,  0.0000000e+00,  0.0000000e+00, ...,
-7.4556941e-05, -1.4621433e-05, -5.7861507e-05], dtype=float32),
'path': '/root/.cache/huggingface/datasets/downloads/extracted/05be0c29807a73c9b099873d2f5975dae6d05e9f7d577458a2466ecb9a2b0c6b/cv-corpus-6.1-2020-12-11/tr/clips/common_voice_tr_21921195.mp3',
'sampling_rate': 16000}


Just like text datasets, you can apply a preprocessing function over an entire dataset with, which is useful for preprocessing all of your audio data at once. Start with a speech recognition model of your choice, and load a processor object that contains:

  1. A feature extractor to convert the speech signal to the model’s input format. Every speech recognition model on the 🤗 Hub contains a predefined feature extractor that can be easily loaded with AutoFeatureExtractor.from_pretrained(...).

  2. A tokenizer to convert the model’s output format to text. Fine-tuned speech recognition models, such as facebook/wav2vec2-base-960h, contain a predefined tokenizer that can be easily loaded with AutoTokenizer.from_pretrained(...).

    For pretrained speech recognition models, such as facebook/wav2vec2-large-xlsr-53, a tokenizer needs to be created from the target text as explained here. The following example demonstrates how to load a feature extractor, tokenizer and processor for a pretrained speech recognition model:

>>> from transformers import AutoTokenizer, AutoFeatureExtractor, Wav2Vec2Processor
>>> model_checkpoint = "facebook/wav2vec2-large-xlsr-53"
>>> # after defining a vocab.json file you can instantiate a tokenizer object:
>>> tokenizer = AutoTokenizer("./vocab.json", unk_token="[UNK]", pad_token="[PAD]", word_delimiter_token="|")
>>> feature_extractor = AutoFeatureExtractor.from_pretrained(model_checkpoint)
>>> processor = Wav2Vec2Processor.from_pretrained(feature_extractor=feature_extractor, tokenizer=tokenizer)

For fine-tuned speech recognition models, you can simply load a predefined processor object with:

>>> from transformers import Wav2Vec2Processor
>>> processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")

Make sure to include the audio key in your preprocessing function when you call so that you are actually resampling the audio data:

>>> def prepare_dataset(batch):
...     audio = batch["audio"]
...     batch["input_values"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
...     batch["input_length"] = len(batch["input_values"])
...     with processor.as_target_processor():
...         batch["labels"] = processor(batch["sentence"]).input_ids
...     return batch
>>> common_voice_train =, remove_columns=common_voice_train.column_names)