Transformers documentation

Automatic speech recognition

You are viewing v4.27.2 version. A newer version v4.42.0 is available.
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Automatic speech recognition

Automatic speech recognition (ASR) converts a speech signal to text, mapping a sequence of audio inputs to text outputs. Virtual assistants like Siri and Alexa use ASR models to help users everyday, and there are many other useful user-facing applications like live captioning and note-taking during meetings.

This guide will show you how to:

  1. Finetune Wav2Vec2 on the MInDS-14 dataset to transcribe audio to text.
  2. Use your finetuned model for inference.
The task illustrated in this tutorial is supported by the following model architectures:

Data2VecAudio, Hubert, M-CTC-T, SEW, SEW-D, UniSpeech, UniSpeechSat, Wav2Vec2, Wav2Vec2-Conformer, WavLM

Before you begin, make sure you have all the necessary libraries installed:

pip install transformers datasets evaluate jiwer

We encourage you to login to your Hugging Face account so you can upload and share your model with the community. When prompted, enter your token to login:

>>> from huggingface_hub import notebook_login

>>> notebook_login()

Load MInDS-14 dataset

Start by loading a smaller subset of the MInDS-14 dataset from the 🤗 Datasets library. This’ll give you a chance to experiment and make sure everything works before spending more time training on the full dataset.

>>> from datasets import load_dataset, Audio

>>> minds = load_dataset("PolyAI/minds14", name="en-US", split="train[:100]")

Split the dataset’s train split into a train and test set with the ~Dataset.train_test_split method:

>>> minds = minds.train_test_split(test_size=0.2)

Then take a look at the dataset:

>>> minds
    train: Dataset({
        features: ['path', 'audio', 'transcription', 'english_transcription', 'intent_class', 'lang_id'],
        num_rows: 16
    test: Dataset({
        features: ['path', 'audio', 'transcription', 'english_transcription', 'intent_class', 'lang_id'],
        num_rows: 4

While the dataset contains a lot of useful information, like lang_id and english_transcription, you’ll focus on the audio and transcription in this guide. Remove the other columns with the remove_columns method:

>>> minds = minds.remove_columns(["english_transcription", "intent_class", "lang_id"])

Take a look at the example again:

>>> minds["train"][0]
{'audio': {'array': array([-0.00024414,  0.        ,  0.        , ...,  0.00024414,
          0.00024414,  0.00024414], dtype=float32),
  'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~APP_ERROR/602ba9e2963e11ccd901cd4f.wav',
  'sampling_rate': 8000},
 'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~APP_ERROR/602ba9e2963e11ccd901cd4f.wav',
 'transcription': "hi I'm trying to use the banking app on my phone and currently my checking and savings account balance is not refreshing"}

There are two fields:

  • audio: a 1-dimensional array of the speech signal that must be called to load and resample the audio file.
  • transcription: the target text.


The next step is to load a Wav2Vec2 processor to process the audio signal:

>>> from transformers import AutoProcessor

>>> processor = AutoProcessor.from_pretrained("facebook/wav2vec2-base")

The MInDS-14 dataset has a sampling rate of 8000kHz (you can find this information in its dataset card), which means you’ll need to resample the dataset to 16000kHz to use the pretrained Wav2Vec2 model:

>>> minds = minds.cast_column("audio", Audio(sampling_rate=16_000))
>>> minds["train"][0]
{'audio': {'array': array([-2.38064706e-04, -1.58618059e-04, -5.43987835e-06, ...,
          2.78103951e-04,  2.38446111e-04,  1.18740834e-04], dtype=float32),
  'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~APP_ERROR/602ba9e2963e11ccd901cd4f.wav',
  'sampling_rate': 16000},
 'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~APP_ERROR/602ba9e2963e11ccd901cd4f.wav',
 'transcription': "hi I'm trying to use the banking app on my phone and currently my checking and savings account balance is not refreshing"}

As you can see in the transcription above, the text contains a mix of upper and lowercase characters. The Wav2Vec2 tokenizer is only trained on uppercase characters so you’ll need to make sure the text matches the tokenizer’s vocabulary:

>>> def uppercase(example):
...     return {"transcription": example["transcription"].upper()}

>>> minds =

Now create a preprocessing function that:

  1. Calls the audio column to load and resample the audio file.
  2. Extracts the input_values from the audio file and tokenize the transcription column with the processor.
>>> def prepare_dataset(batch):
...     audio = batch["audio"]
...     batch = processor(audio["array"], sampling_rate=audio["sampling_rate"], text=batch["transcription"])
...     batch["input_length"] = len(batch["input_values"][0])
...     return batch

To apply the preprocessing function over the entire dataset, use 🤗 Datasets map function. You can speed up map by increasing the number of processes with the num_proc parameter. Remove the columns you don’t need with the remove_columns method:

>>> encoded_minds =, remove_columns=minds.column_names["train"], num_proc=4)

🤗 Transformers doesn’t have a data collator for ASR, so you’ll need to adapt the DataCollatorWithPadding to create a batch of examples. It’ll also dynamically pad your text and labels to the length of the longest element in its batch (instead of the entire dataset) so they are a uniform length. While it is possible to pad your text in the tokenizer function by setting padding=True, dynamic padding is more efficient.

Unlike other data collators, this specific data collator needs to apply a different padding method to input_values and labels:

>>> import torch

>>> from dataclasses import dataclass, field
>>> from typing import Any, Dict, List, Optional, Union

>>> @dataclass
... class DataCollatorCTCWithPadding:
...     processor: AutoProcessor
...     padding: Union[bool, str] = "longest"

...     def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
...         # split inputs and labels since they have to be of different lengths and need
...         # different padding methods
...         input_features = [{"input_values": feature["input_values"][0]} for feature in features]
...         label_features = [{"input_ids": feature["labels"]} for feature in features]

...         batch = self.processor.pad(input_features, padding=self.padding, return_tensors="pt")

...         labels_batch = self.processor.pad(labels=label_features, padding=self.padding, return_tensors="pt")

...         # replace padding with -100 to ignore loss correctly
...         labels = labels_batch["input_ids"].masked_fill(, -100)

...         batch["labels"] = labels

...         return batch

Now instantiate your DataCollatorForCTCWithPadding:

>>> data_collator = DataCollatorCTCWithPadding(processor=processor, padding="longest")


Including a metric during training is often helpful for evaluating your model’s performance. You can quickly load a evaluation method with the 🤗 Evaluate library. For this task, load the word error rate (WER) metric (see the 🤗 Evaluate quick tour to learn more about how to load and compute a metric):

>>> import evaluate

>>> wer = evaluate.load("wer")

Then create a function that passes your predictions and labels to compute to calculate the WER:

>>> import numpy as np

>>> def compute_metrics(pred):
...     pred_logits = pred.predictions
...     pred_ids = np.argmax(pred_logits, axis=-1)

...     pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id

...     pred_str = processor.batch_decode(pred_ids)
...     label_str = processor.batch_decode(pred.label_ids, group_tokens=False)

...     wer = wer.compute(predictions=pred_str, references=label_str)

...     return {"wer": wer}

Your compute_metrics function is ready to go now, and you’ll return to it when you setup your training.


Hide Pytorch content

If you aren’t familiar with finetuning a model with the Trainer, take a look at the basic tutorial here!

You’re ready to start training your model now! Load Wav2Vec2 with AutoModelForCTC. Specify the reduction to apply with the ctc_loss_reduction parameter. It is often better to use the average instead of the default summation:

>>> from transformers import AutoModelForCTC, TrainingArguments, Trainer

>>> model = AutoModelForCTC.from_pretrained(
...     "facebook/wav2vec2-base",
...     ctc_loss_reduction="mean",
...     pad_token_id=processor.tokenizer.pad_token_id,
... )

At this point, only three steps remain:

  1. Define your training hyperparameters in TrainingArguments. The only required parameter is output_dir which specifies where to save your model. You’ll push this model to the Hub by setting push_to_hub=True (you need to be signed in to Hugging Face to upload your model). At the end of each epoch, the Trainer will evaluate the WER and save the training checkpoint.
  2. Pass the training arguments to Trainer along with the model, dataset, tokenizer, data collator, and compute_metrics function.
  3. Call train() to finetune your model.
>>> training_args = TrainingArguments(
...     output_dir="my_awesome_asr_mind_model",
...     per_device_train_batch_size=8,
...     gradient_accumulation_steps=2,
...     learning_rate=1e-5,
...     warmup_steps=500,
...     max_steps=2000,
...     gradient_checkpointing=True,
...     fp16=True,
...     group_by_length=True,
...     evaluation_strategy="steps",
...     per_device_eval_batch_size=8,
...     save_steps=1000,
...     eval_steps=1000,
...     logging_steps=25,
...     load_best_model_at_end=True,
...     metric_for_best_model="wer",
...     greater_is_better=False,
...     push_to_hub=True,
... )

>>> trainer = Trainer(
...     model=model,
...     args=training_args,
...     train_dataset=encoded_minds["train"],
...     eval_dataset=encoded_minds["test"],
...     tokenizer=processor.feature_extractor,
...     data_collator=data_collator,
...     compute_metrics=compute_metrics,
... )

>>> trainer.train()

Once training is completed, share your model to the Hub with the push_to_hub() method so everyone can use your model:

>>> trainer.push_to_hub()

For a more in-depth example of how to finetune a model for automatic speech recognition, take a look at this blog post for English ASR and this post for multilingual ASR.


Great, now that you’ve finetuned a model, you can use it for inference!

Load an audio file you’d like to run inference on. Remember to resample the sampling rate of the audio file to match the sampling rate of the model if you need to!

>>> from datasets import load_dataset, Audio

>>> dataset = load_dataset("PolyAI/minds14", "en-US", split="train")
>>> dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
>>> sampling_rate = dataset.features["audio"].sampling_rate
>>> audio_file = dataset[0]["audio"]["path"]

The simplest way to try out your finetuned model for inference is to use it in a pipeline(). Instantiate a pipeline for automatic speech recognition with your model, and pass your audio file to it:

>>> from transformers import pipeline

>>> transcriber = pipeline("automatic-speech-recognition", model="stevhliu/my_awesome_asr_minds_model")
>>> transcriber(audio_file)

The transcription is decent, but it could be better! Try finetuning your model on more examples to get even better results!

You can also manually replicate the results of the pipeline if you’d like:

Hide Pytorch content

Load a processor to preprocess the audio file and transcription and return the input as PyTorch tensors:

>>> from transformers import AutoProcessor

>>> processor = AutoProcessor.from_pretrained("stevhliu/my_awesome_asr_mind_model")
>>> inputs = processor(dataset[0]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt")

Pass your inputs to the model and return the logits:

>>> from transformers import AutoModelForCTC

>>> model = AutoModelForCTC.from_pretrained("stevhliu/my_awesome_asr_mind_model")
>>> with torch.no_grad():
...     logits = model(**inputs).logits

Get the predicted input_ids with the highest probability, and use the processor to decode the predicted input_ids back into text:

>>> import torch

>>> predicted_ids = torch.argmax(logits, dim=-1)
>>> transcription = processor.batch_decode(predicted_ids)
>>> transcription