Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers

In this blog, we present a step-by-step guide on fine-tuning Whisper for any multilingual ASR dataset using Hugging Face 🤗 Transformers. This blog provides in-depth explanations of the Whisper model, the Common Voice dataset and the theory behind fine-tuning, with accompanying code cells to execute the data preparation and fine-tuning steps. For a more streamlined version of the notebook with fewer explanations but all the code, see the accompanying Google Colab.

Table of Contents

1. Introduction
2. Fine-tuning Whisper in a Google Colab
   1. Prepare Environment
   2. Load Dataset
   3. Prepare Feature Extractor, Tokenizer and Data
   4. Training and Evaluation
   5. Building a Demo
3. Closing Remarks

Introduction

Whisper is a pre-trained model for automatic speech recognition (ASR) published in September 2022 by the authors Alec Radford et al. from OpenAI. Unlike many of its predecessors, such as Wav2Vec 2.0, which are pre-trained on un-labelled audio data, Whisper is pre-trained on a vast quantity of labelled audio-transcription data, 680,000 hours to be precise. This is an order of magnitude more data than the un-labelled audio data used to train Wav2Vec 2.0 (60,000 hours). What is more, 117,000 hours of this pre-training data is multilingual ASR data. This results in checkpoints that can be applied to over 96 languages, many of which are considered low-resource.

This quantity of labelled data enables Whisper to be pre-trained directly on the supervised task of speech recognition, learning a speech-to-text mapping from the labelled audio-transcription pre-training data ${}^1$. As a consequence, Whisper requires little additional fine-tuning to yield a performant ASR model. This is in contrast to Wav2Vec 2.0, which is pre-trained on the unsupervised task of masked prediction. Here, the model is trained to learn an intermediate mapping from speech to hidden states from un-labelled audio only data. While unsupervised pre-training yields high-quality representations of speech, it does not learn a speech-to-text mapping. This mapping is only learned during fine-tuning, thus requiring more fine-tuning to yield competitive performance.

When scaled to 680,000 hours of labelled pre-training data, Whisper models demonstrate a strong ability to generalise to many datasets and domains. The pre-trained checkpoints achieve competitive results to state-of-the-art ASR systems, with near 3% word error rate (WER) on the test-clean subset of LibriSpeech ASR and a new state-of-the-art on TED-LIUM with 4.7% WER (c.f. Table 8 of the Whisper paper). The extensive multilingual ASR knowledge acquired by Whisper during pre-training can be leveraged for other low-resource languages; through fine-tuning, the pre-trained checkpoints can be adapted for specific datasets and languages to further improve upon these results.

Whisper is a Transformer based encoder-decoder model, also referred to as a sequence-to-sequence model. It maps a sequence of audio spectrogram features to a sequence of text tokens. First, the raw audio inputs are converted to a log-Mel spectrogram by action of the feature extractor. The Transformer encoder then encodes the spectrogram to form a sequence of encoder hidden states. Finally, the decoder autoregressively predicts text tokens, conditional on both the previous tokens and the encoder hidden states. Figure 1 summarises the Whisper model.

In a sequence-to-sequence model, the encoder transforms the audio inputs into a set of hidden state representations, extracting important features from the spoken speech. The decoder plays the role of a language model, processing the hidden state representations and generating the corresponding text transcriptions. Incorporating a language model internally in the system architecture is termed deep fusion. This is in contrast to shallow fusion, where a language model is combined externally with an encoder, such as with CTC + $n$-gram (c.f. Internal Language Model Estimation). With deep fusion, the entire system can be trained end-to-end with the same training data and loss function, giving greater flexibility and generally superior performance (c.f. ESB Benchmark).

Whisper is pre-trained and fine-tuned using the cross-entropy objective function, a standard objective function for training sequence-to-sequence systems on classification tasks. Here, the system is trained to correctly classify the target text token from a pre-defined vocabulary of text tokens.

The Whisper checkpoints come in five configurations of varying model sizes. The smallest four are trained on either English-only or multilingual data. The largest checkpoints are multilingual only. All 11 of the pre-trained checkpoints are available on the Hugging Face Hub. The checkpoints are summarised in the following table with links to the models on the Hub:

For demonstration purposes, we'll fine-tune the multilingual version of the small checkpoint with 244M params (~= 1GB). As for our data, we'll train and evaluate our system on a low-resource language taken from the Common Voice dataset. We'll show that with as little as 8 hours of fine-tuning data, we can achieve strong performance in this language.

---

${}^1$ The name Whisper follows from the acronym “WSPSR”, which stands for “Web-scale Supervised Pre-training for Speech Recognition”.

Fine-tuning Whisper in a Google Colab

Prepare Environment

We'll employ several popular Python packages to fine-tune the Whisper model. We'll use datasets[audio] to download and prepare our training data, alongside transformers and accelerate to load and train our Whisper model. We'll also require the soundfile package to pre-process audio files, evaluate and jiwer to assess the performance of our model, and tensorboard to log our metrics. Finally, we'll use gradio to build a flashy demo of our fine-tuned model.

!pip install --upgrade pip
!pip install --upgrade datasets[audio] transformers accelerate evaluate jiwer tensorboard gradio

We strongly advise you to upload model checkpoints directly the Hugging Face Hub whilst training. The Hub provides:

• Integrated version control: you can be sure that no model checkpoint is lost during training.
• Tensorboard logs: track important metrics over the course of training.
• Model cards: document what a model does and its intended use cases.
• Community: an easy way to share and collaborate with the community!

Linking the notebook to the Hub is straightforward - it simply requires entering your Hub authentication token when prompted. Find your Hub authentication token here:

from huggingface_hub import notebook_login

notebook_login()

Print Output:

Login successful
Your token has been saved to /root/.huggingface/token

Load Dataset

Common Voice is a series of crowd-sourced datasets where speakers record text from Wikipedia in various languages. We'll use the latest edition of the Common Voice dataset at the time of writing (version 11). As for our language, we'll fine-tune our model on Hindi, an Indo-Aryan language spoken in northern, central, eastern, and western India. Common Voice 11.0 contains approximately 12 hours of labelled Hindi data, 4 of which are held-out test data.

Tip: you can find the latest version of the Common Voice dataset by checking the Mozilla Foundation organisation page on the Hugging Face Hub. Later versions cover more languages and contain more data per-language.

Let's head to the Hub and view the dataset page for Common Voice: mozilla-foundation/common_voice_11_0.

The first time we view this page, we'll be asked to accept the terms of use. After that, we'll be given full access to the dataset.

Once we've provided authentication to use the dataset, we'll be presented with the dataset preview. The dataset preview shows us the first 100 samples of the dataset. What's more, it's loaded up with audio samples ready for us to listen to in real time. We can select the Hindi subset of Common Voice by setting the subset to hi using the dropdown menu (hi being the language identifier code for Hindi):

If we hit the play button on the first sample, we can listen to the audio and see the corresponding text. Have a scroll through the samples for the train and test sets to get a better feel for the audio and text data that we're dealing with. You can tell from the intonation and style that the recordings are taken from narrated speech. You'll also likely notice the large variation in speakers and recording quality, a common trait of crowd-sourced data.

Using 🤗 Datasets, downloading and preparing data is extremely simple. We can download and prepare the Common Voice splits in just one line of code. Since Hindi is very low-resource, we'll combine the train and validation splits to give approximately 8 hours of training data. We'll use the 4 hours of test data as our held-out test set:

from datasets import load_dataset, DatasetDict

common_voice = DatasetDict()

common_voice["train"] = load_dataset("mozilla-foundation/common_voice_11_0", "hi", split="train+validation", use_auth_token=True)
common_voice["test"] = load_dataset("mozilla-foundation/common_voice_11_0", "hi", split="test", use_auth_token=True)

print(common_voice)

Print Output:

DatasetDict({
    train: Dataset({
        features: ['client_id', 'path', 'audio', 'sentence', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'],
        num_rows: 6540
    })
    test: Dataset({
        features: ['client_id', 'path', 'audio', 'sentence', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'],
        num_rows: 2894
    })
})

Most ASR datasets only provide input audio samples (audio) and the corresponding transcribed text (sentence). Common Voice contains additional metadata information, such as accent and locale, which we can disregard for ASR. Keeping the notebook as general as possible, we only consider the input audio and transcribed text for fine-tuning, discarding the additional metadata information:

common_voice = common_voice.remove_columns(["accent", "age", "client_id", "down_votes", "gender", "locale", "path", "segment", "up_votes"])

Common Voice is but one multilingual ASR dataset that we can download from the Hub - there are plenty more available to us! To view the range of datasets available for speech recognition, follow the link: ASR Datasets on the Hub.

Prepare Feature Extractor, Tokenizer and Data

The ASR pipeline can be de-composed into three components:

1. A feature extractor which pre-processes the raw audio-inputs
2. The model which performs the sequence-to-sequence mapping
3. A tokenizer which post-processes the model outputs to text format

In 🤗 Transformers, the Whisper model has an associated feature extractor and tokenizer, called WhisperFeatureExtractor and WhisperTokenizer respectively.

We'll go through details of the feature extractor and tokenizer one-by-one!

Load WhisperFeatureExtractor

Speech is represented by a 1-dimensional array that varies with time. The value of the array at any given time step is the signal's amplitude at that point. From the amplitude information alone, we can reconstruct the frequency spectrum of the audio and recover all acoustic features.

Since speech is continuous, it contains an infinite number of amplitude values. This poses problems for computer devices which expect finite arrays. Thus, we discretise our speech signal by sampling values from our signal at fixed time steps. The interval with which we sample our audio is known as the sampling rate and is usually measured in samples/sec or Hertz (Hz). Sampling with a higher sampling rate results in a better approximation of the continuous speech signal, but also requires storing more values per second.

It is crucial that we match the sampling rate of our audio inputs to the sampling rate expected by our model, as audio signals with different sampling rates have very different distributions. Audio samples should only ever be processed with the correct sampling rate. Failing to do so can lead to unexpected results! For instance, taking an audio sample with a sampling rate of 16kHz and listening to it with a sampling rate of 8kHz will make the audio sound as though it's in half-speed. In the same way, passing audio with the wrong sampling rate can falter an ASR model that expects one sampling rate and receives another. The Whisper feature extractor expects audio inputs with a sampling rate of 16kHz, so we need to match our inputs to this value. We don't want to inadvertently train an ASR system on slow-motion speech!

The Whisper feature extractor performs two operations. It first pads/truncates a batch of audio samples such that all samples have an input length of 30s. Samples shorter than 30s are padded to 30s by appending zeros to the end of the sequence (zeros in an audio signal corresponding to no signal or silence). Samples longer than 30s are truncated to 30s. Since all elements in the batch are padded/truncated to a maximum length in the input space, we don't require an attention mask when forwarding the audio inputs to the Whisper model. Whisper is unique in this regard - with most audio models, you can expect to provide an attention mask that details where sequences have been padded, and thus where they should be ignored in the self-attention mechanism. Whisper is trained to operate without an attention mask and infer directly from the speech signals where to ignore the inputs.

The second operation that the Whisper feature extractor performs is converting the padded audio arrays to log-Mel spectrograms. These spectrograms are a visual representation of the frequencies of a signal, rather like a Fourier transform. An example spectrogram is shown in Figure 2. Along the $y$-axis are the Mel channels, which correspond to particular frequency bins. Along the $x$-axis is time. The colour of each pixel corresponds to the log-intensity of that frequency bin at a given time. The log-Mel spectrogram is the form of input expected by the Whisper model.

The Mel channels (frequency bins) are standard in speech processing and chosen to approximate the human auditory range. All we need to know for Whisper fine-tuning is that the spectrogram is a visual representation of the frequencies in the speech signal. For more detail on Mel channels, refer to Mel-frequency cepstrum.

Luckily for us, the 🤗 Transformers Whisper feature extractor performs both the padding and spectrogram conversion in just one line of code! Let's go ahead and load the feature extractor from the pre-trained checkpoint to have ready for our audio data:

from transformers import WhisperFeatureExtractor

feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-small")

Load WhisperTokenizer

Now let's look at how to load a Whisper tokenizer. The Whisper model outputs text tokens that indicate the index of the predicted text among the dictionary of vocabulary items. The tokenizer maps a sequence of text tokens to the actual text string (e.g. [1169, 3797, 3332] -> "the cat sat").

Traditionally, when using encoder-only models for ASR, we decode using Connectionist Temporal Classification (CTC). Here we are required to train a CTC tokenizer for each dataset we use. One of the advantages of using an encoder-decoder architecture is that we can directly leverage the tokenizer from the pre-trained model.

The Whisper tokenizer is pre-trained on the transcriptions for the 96 pre-training languages. Consequently, it has an extensive byte-pair that is appropriate for almost all multilingual ASR applications. For Hindi, we can load the tokenizer and use it for fine-tuning without any further modifications. We simply have to specify the target language and the task. These arguments inform the tokenizer to prefix the language and task tokens to the start of encoded label sequences:

from transformers import WhisperTokenizer

tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-small", language="Hindi", task="transcribe")

Tip: the blog post can be adapted for speech translation by setting the task to "translate" and the language to the target text language in the above line. This will prepend the relevant task and language tokens for speech translation when the dataset is pre-processed.

We can verify that the tokenizer correctly encodes Hindi characters by encoding and decoding the first sample of the Common Voice dataset. When encoding the transcriptions, the tokenizer appends 'special tokens' to the start and end of the sequence, including the start/end of transcript tokens, the language token and the task tokens (as specified by the arguments in the previous step). When decoding the label ids, we have the option of 'skipping' these special tokens, allowing us to return a string in the original input form:

input_str = common_voice["train"][0]["sentence"]
labels = tokenizer(input_str).input_ids
decoded_with_special = tokenizer.decode(labels, skip_special_tokens=False)
decoded_str = tokenizer.decode(labels, skip_special_tokens=True)

print(f"Input:                 {input_str}")
print(f"Decoded w/ special:    {decoded_with_special}")
print(f"Decoded w/out special: {decoded_str}")
print(f"Are equal:             {input_str == decoded_str}")

Print Output:

Input:                 खीर की मिठास पर गरमाई बिहार की सियासत, कुशवाहा ने दी सफाई
Decoded w/ special:    <|startoftranscript|><|hi|><|transcribe|><|notimestamps|>खीर की मिठास पर गरमाई बिहार की सियासत, कुशवाहा ने दी सफाई<|endoftext|>
Decoded w/out special: खीर की मिठास पर गरमाई बिहार की सियासत, कुशवाहा ने दी सफाई
Are equal:             True

Combine To Create A WhisperProcessor

To simplify using the feature extractor and tokenizer, we can wrap both into a single WhisperProcessor class. This processor object inherits from the WhisperFeatureExtractor and WhisperProcessor and can be used on the audio inputs and model predictions as required. In doing so, we only need to keep track of two objects during training: the processor and the model:

from transformers import WhisperProcessor

processor = WhisperProcessor.from_pretrained("openai/whisper-small", language="Hindi", task="transcribe")

Prepare Data

Let's print the first example of the Common Voice dataset to see what form the data is in:

print(common_voice["train"][0])

Print Output:

{'audio': {'path': '/home/sanchit_huggingface_co/.cache/huggingface/datasets/downloads/extracted/607848c7e74a89a3b5225c0fa5ffb9470e39b7f11112db614962076a847f3abf/cv-corpus-11.0-2022-09-21/hi/clips/common_voice_hi_25998259.mp3', 
           'array': array([0.0000000e+00, 0.0000000e+00, 0.0000000e+00, ..., 9.6724887e-07,
       1.5334779e-06, 1.0415988e-06], dtype=float32), 
           'sampling_rate': 48000},
 'sentence': 'खीर की मिठास पर गरमाई बिहार की सियासत, कुशवाहा ने दी सफाई'}

We can see that we've got a 1-dimensional input audio array and the corresponding target transcription. We've spoken heavily about the importance of the sampling rate and the fact that we need to match the sampling rate of our audio to that of the Whisper model (16kHz). Since our input audio is sampled at 48kHz, we need to downsample it to 16kHz before passing it to the Whisper feature extractor.

We'll set the audio inputs to the correct sampling rate using dataset's cast_column method. This operation does not change the audio in-place, but rather signals to datasets to resample audio samples on the fly the first time that they are loaded:

from datasets import Audio

common_voice = common_voice.cast_column("audio", Audio(sampling_rate=16000))

Re-loading the first audio sample in the Common Voice dataset will resample it to the desired sampling rate:

print(common_voice["train"][0])

Print Output:

{'audio': {'path': '/home/sanchit_huggingface_co/.cache/huggingface/datasets/downloads/extracted/607848c7e74a89a3b5225c0fa5ffb9470e39b7f11112db614962076a847f3abf/cv-corpus-11.0-2022-09-21/hi/clips/common_voice_hi_25998259.mp3', 
           'array': array([ 0.0000000e+00,  0.0000000e+00,  0.0000000e+00, ...,
       -3.4206650e-07,  3.2979898e-07,  1.0042874e-06], dtype=float32),
           'sampling_rate': 16000},
 'sentence': 'खीर की मिठास पर गरमाई बिहार की सियासत, कुशवाहा ने दी सफाई'}

Great! We can see that the sampling rate has been downsampled to 16kHz. The array values are also different, as we've now only got approximately one amplitude value for every three we had before.

Now we can write a function to prepare our data ready for the model:

1. We load and resample the audio data by calling batch["audio"]. As explained above, 🤗 Datasets performs any necessary resampling operations on the fly.
2. We use the feature extractor to compute the log-Mel spectrogram input features from our 1-dimensional audio array.
3. We encode the transcriptions to label ids through the use of the tokenizer.

def prepare_dataset(batch):
    # load and resample audio data from 48 to 16kHz
    audio = batch["audio"]

    # compute log-Mel input features from input audio array 
    batch["input_features"] = feature_extractor(audio["array"], sampling_rate=audio["sampling_rate"]).input_features[0]

    # encode target text to label ids 
    batch["labels"] = tokenizer(batch["sentence"]).input_ids
    return batch

We can apply the data preparation function to all of our training examples using dataset's .map method:

common_voice = common_voice.map(prepare_dataset, remove_columns=common_voice.column_names["train"], num_proc=4)

Alright! With that we have our data fully prepared for training! Let's continue and take a look at how we can use this data to fine-tune Whisper.

Note: Currently datasets makes use of both torchaudio and librosa for audio loading and resampling. If you wish to implement your own customised data loading/sampling, you can use the "path" column to obtain the audio file path and disregard the "audio" column.

Training and Evaluation

Now that we've prepared our data, we're ready to dive into the training pipeline. The 🤗 Trainer will do much of the heavy lifting for us. All we have to do is:

• Load a pre-trained checkpoint: we need to load a pre-trained checkpoint and configure it correctly for training.

• Define a data collator: the data collator takes our pre-processed data and prepares PyTorch tensors ready for the model.

• Evaluation metrics: during evaluation, we want to evaluate the model using the word error rate (WER) metric. We need to define a compute_metrics function that handles this computation.

• Define the training arguments: these will be used by the 🤗 Trainer in constructing the training schedule.

Once we've fine-tuned the model, we will evaluate it on the test data to verify that we have correctly trained it to transcribe speech in Hindi.

Load a Pre-Trained Checkpoint

We'll start our fine-tuning run from the pre-trained Whisper small checkpoint. To do this, we'll load the pre-trained weights from the Hugging Face Hub. Again, this is trivial through use of 🤗 Transformers!

from transformers import WhisperForConditionalGeneration

model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

At inference time, the Whisper model automatically detects the language of the source audio and predicts token ids in this language. In cases where the source audio language is known a-priori, such as multilingual fine-tuning, it is beneficial to set the language explicitly. This negates the scenarios when the incorrect language is predicted, causing the predicted text to diverge from the true language during generation. To do so, we set the langauge and task arguments to the generation config. We'll also set any forced_decoder_ids to None, since this was the legacy way of setting the language and task arguments:

model.generation_config.language = "hindi"
model.generation_config.task = "transcribe"

model.generation_config.forced_decoder_ids = None

Define a Data Collator

The data collator for a sequence-to-sequence speech model is unique in the sense that it treats the input_features and labels independently: the input_features must be handled by the feature extractor and the labels by the tokenizer.

The input_features are already padded to 30s and converted to a log-Mel spectrogram of fixed dimension, so all we have to do is convert them to batched PyTorch tensors. We do this using the feature extractor's .pad method with return_tensors=pt. Note that no additional padding is applied here since the inputs are of fixed dimension, the input_features are simply converted to PyTorch tensors.

On the other hand, the labels are un-padded. We first pad the sequences to the maximum length in the batch using the tokenizer's .pad method. The padding tokens are then replaced by -100 so that these tokens are not taken into account when computing the loss. We then cut the start of transcript token from the beginning of the label sequence as we append it later during training.

We can leverage the WhisperProcessor we defined earlier to perform both the feature extractor and the tokenizer operations:

import torch

from dataclasses import dataclass
from typing import Any, Dict, List, Union

@dataclass
class DataCollatorSpeechSeq2SeqWithPadding:
    processor: Any
    decoder_start_token_id: int

    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
        # first treat the audio inputs by simply returning torch tensors
        input_features = [{"input_features": feature["input_features"]} for feature in features]
        batch = self.processor.feature_extractor.pad(input_features, return_tensors="pt")

        # get the tokenized label sequences
        label_features = [{"input_ids": feature["labels"]} for feature in features]
        # pad the labels to max length
        labels_batch = self.processor.tokenizer.pad(label_features, return_tensors="pt")

        # replace padding with -100 to ignore loss correctly
        labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)

        # if bos token is appended in previous tokenization step,
        # cut bos token here as it's append later anyways
        if (labels[:, 0] == self.decoder_start_token_id).all().cpu().item():
            labels = labels[:, 1:]

        batch["labels"] = labels

        return batch

Let's initialise the data collator we've just defined:

data_collator = DataCollatorSpeechSeq2SeqWithPadding(
    processor=processor,
    decoder_start_token_id=model.config.decoder_start_token_id,
)

Evaluation Metrics

Next, we define the evaluation metric we'll use on our evaluation set. We'll use the Word Error Rate (WER) metric, the 'de-facto' metric for assessing ASR systems. For more information, refer to the WER docs. We'll load the WER metric from 🤗 Evaluate:

import evaluate

metric = evaluate.load("wer")

We then simply have to define a function that takes our model predictions and returns the WER metric. This function, called compute_metrics, first replaces -100 with the pad_token_id in the label_ids (undoing the step we applied in the data collator to ignore padded tokens correctly in the loss). It then decodes the predicted and label ids to strings. Finally, it computes the WER between the predictions and reference labels:

def compute_metrics(pred):
    pred_ids = pred.predictions
    label_ids = pred.label_ids

    # replace -100 with the pad_token_id
    label_ids[label_ids == -100] = tokenizer.pad_token_id

    # we do not want to group tokens when computing the metrics
    pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
    label_str = tokenizer.batch_decode(label_ids, skip_special_tokens=True)

    wer = 100 * metric.compute(predictions=pred_str, references=label_str)

    return {"wer": wer}

Define the Training Arguments

In the final step, we define all the parameters related to training. A subset of parameters are explained below:

• output_dir: local directory in which to save the model weights. This will also be the repository name on the Hugging Face Hub.
• generation_max_length: maximum number of tokens to autoregressively generate during evaluation.
• save_steps: during training, intermediate checkpoints will be saved and uploaded asynchronously to the Hub every save_steps training steps.
• eval_steps: during training, evaluation of intermediate checkpoints will be performed every eval_steps training steps.
• report_to: where to save training logs. Supported platforms are "azure_ml", "comet_ml", "mlflow", "neptune", "tensorboard" and "wandb". Pick your favourite or leave as "tensorboard" to log to the Hub.

For more detail on the other training arguments, refer to the Seq2SeqTrainingArguments docs.

from transformers import Seq2SeqTrainingArguments

training_args = Seq2SeqTrainingArguments(
    output_dir="./whisper-small-hi",  # change to a repo name of your choice
    per_device_train_batch_size=16,
    gradient_accumulation_steps=1,  # increase by 2x for every 2x decrease in batch size
    learning_rate=1e-5,
    warmup_steps=500,
    max_steps=5000,
    gradient_checkpointing=True,
    fp16=True,
    evaluation_strategy="steps",
    per_device_eval_batch_size=8,
    predict_with_generate=True,
    generation_max_length=225,
    save_steps=1000,
    eval_steps=1000,
    logging_steps=25,
    report_to=["tensorboard"],
    load_best_model_at_end=True,
    metric_for_best_model="wer",
    greater_is_better=False,
    push_to_hub=True,
)

Note: if one does not want to upload the model checkpoints to the Hub, set push_to_hub=False.

We can forward the training arguments to the 🤗 Trainer along with our model, dataset, data collator and compute_metrics function:

from transformers import Seq2SeqTrainer

trainer = Seq2SeqTrainer(
    args=training_args,
    model=model,
    train_dataset=common_voice["train"],
    eval_dataset=common_voice["test"],
    data_collator=data_collator,
    compute_metrics=compute_metrics,
    tokenizer=processor.feature_extractor,
)

And with that, we're ready to start training!

Training

To launch training, simply execute:

trainer.train()

Training will take approximately 5-10 hours depending on your GPU or the one allocated to the Google Colab. Depending on your GPU, it is possible that you will encounter a CUDA "out-of-memory" error when you start training. In this case, you can reduce the per_device_train_batch_size incrementally by factors of 2 and employ gradient_accumulation_steps to compensate.

Print Output:

Our best WER is 32.0% after 4000 training steps. For reference, the pre-trained Whisper small model achieves a WER of 63.5%, meaning we achieve an improvement of 31.5% absolute through fine-tuning. Not bad for just 8h of training data!

We're now ready to share our fine-tuned model on the Hugging Face Hub. To make it more accessible with appropriate tags and README information, we can set the appropriate key-word arguments (kwargs) when we push. You can change these values to match your dataset, language and model name accordingly:

kwargs = {
    "dataset_tags": "mozilla-foundation/common_voice_11_0",
    "dataset": "Common Voice 11.0",  # a 'pretty' name for the training dataset
    "dataset_args": "config: hi, split: test",
    "language": "hi",
    "model_name": "Whisper Small Hi - Sanchit Gandhi",  # a 'pretty' name for your model
    "finetuned_from": "openai/whisper-small",
    "tasks": "automatic-speech-recognition",
}

The training results can now be uploaded to the Hub. To do so, execute the push_to_hub command:

trainer.push_to_hub(**kwargs)

You can now share this model with anyone using the link on the Hub. They can also load it with the identifier "your-username/the-name-you-picked", for instance:

from transformers import WhisperForConditionalGeneration, WhisperProcessor

model = WhisperForConditionalGeneration.from_pretrained("sanchit-gandhi/whisper-small-hi")
processor = WhisperProcessor.from_pretrained("sanchit-gandhi/whisper-small-hi")

While the fine-tuned model yields satisfactory results on the Common Voice Hindi test data, it is by no means optimal. The purpose of this notebook is to demonstrate how the pre-trained Whisper checkpoints can be fine-tuned on any multilingual ASR dataset. The results could likely be improved by optimising the training hyperparameters, such as learning rate and dropout, and using a larger pre-trained checkpoint (medium or large-v3).

Building a Demo

Now that we've fine-tuned our model, we can build a demo to show off its ASR capabilities! We'll use 🤗 Transformers pipeline, which will take care of the entire ASR pipeline, right from pre-processing the audio inputs to decoding the model predictions. We'll build our interactive demo with Gradio. Gradio is arguably the most straightforward way of building machine learning demos; with Gradio, we can build a demo in just a matter of minutes!

Running the example below will generate a Gradio demo where we can record speech through the microphone of our computer and input it to our fine-tuned Whisper model to transcribe the corresponding text:

from transformers import pipeline
import gradio as gr

pipe = pipeline(model="sanchit-gandhi/whisper-small-hi")  # change to "your-username/the-name-you-picked"

def transcribe(audio):
    text = pipe(audio)["text"]
    return text

iface = gr.Interface(
    fn=transcribe, 
    inputs=gr.Audio(source="microphone", type="filepath"), 
    outputs="text",
    title="Whisper Small Hindi",
    description="Realtime demo for Hindi speech recognition using a fine-tuned Whisper small model.",
)

iface.launch()

Closing Remarks

In this blog, we covered a step-by-step guide on fine-tuning Whisper for multilingual ASR using 🤗 Datasets, Transformers and the Hugging Face Hub. Refer to the Google Colab should you wish to try fine-tuning for yourself. If you're interested in fine-tuning other Transformers models, both for English and multilingual ASR, be sure to check out the examples scripts at examples/pytorch/speech-recognition.