nullonesix/training

Training Distil-Whisper

This sub-folder contains all the scripts required to train a Distil-Whisper model in your choice of language. They are slightly modified from the original scripts used to distill Whisper for English ASR (as-per the Distil-Whisper paper). The main difference is that these scripts are written in PyTorch, whereas the original scripts are in JAX/Flax. These scripts are also made to be easier to run end-to-end, whereas the original scripts require more steps and are somewhat hard-coded for English ASR. Both sets of scripts achieve equivalent downstream results when the hyper-parameters are set equal.

If you are interested in reproducing the original Distil-Whisper checkpoints, we refer you to the sub-folder Flax Training. Otherwise, if you wish to distill Whisper on your own language/dataset, we recommend you use these scripts for ease of use and the configurability they provide.

Reproducing the Distil-Whisper project requires four stages to be completed in successive order:

1. Pseudo-labelling
2. Initialisation
3. Training
4. Evaluation

This README is partitioned according to the four stages. Each section provides a minimal example for running the scripts used in the project. We will use a running example of distilling the Whisper model for Hindi speech recognition on the Common Voice dataset. Note that this dataset only contains ~20 hours of audio data. Thus, it can be run extremely quickly, but does not provide sufficient data to achieve optimal performance. We recommend training on upwards of 1000 hours of data should you want to match the performance of Whisper on high-resource languages.

Requirements

The Distil-Whisper training code is written in PyTorch and Accelerate. It heavily leverages the Whisper implementation in 🤗 Transformers for both training and inference.

The instructions for installing the package are as follows:

1. Install PyTorch from the official instructions, ensuring you install the correct version for your hardware and CUDA version.
2. Fork the distil-whisper repository by clicking on the fork button on the reopsitory's page
3. Clone the distil-whisper repository and add the base repository as a remote. This will allow you to "pull" any upstream changes that are made to the base repository:

git clone https://github.com/<your GitHub handle>/distil-whisper.git
cd distil-whisper
git remote add upstream https://github.com/huggingface/distil-whisper.git

4. pip install the required packages from the setup.py file:

cd training
pip install -e .
cd ../..

5. Configure Accelerate by running the following command. Note that you should set the number of GPUs you wish to use for distillation, and also the data type (dtype) to your preferred dtype for training/inference (e.g. bfloat16 on A100 GPUs, float16 on V100 GPUs, etc.):

accelerate config

6. The last thing we need to do is link our Hugging Face account so that we can pull/push model repositories on the Hub. This will allow us to save our final distilled weights on the Hub so that we can share them with the community. Run the command:

git config --global credential.helper store
huggingface-cli login

And then enter an authentication token from https://huggingface.co/settings/tokens. Create a new token if you do not have one already. You should make sure that this token has "write" privileges.

To confirm that you have a working environment, first accept the terms of use of the Common Voice 16.1 dataset on the Hub: https://huggingface.co/datasets/mozilla-foundation/common_voice_16_1

You can run the following code cell to stream one sample of data from the Common Voice dataset, and check that you can perform inference using the "tiny" Whisper model:

from transformers import WhisperProcessor, WhisperForConditionalGeneration
from datasets import load_dataset, Audio

model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny", low_cpu_mem_usage=True)
processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")

model.to("cuda")

common_voice = load_dataset("mozilla-foundation/common_voice_16_1", "en", split="validation", streaming=True)
common_voice = common_voice.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))

inputs = processor(next(iter(common_voice))["audio"]["array"], sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features

generated_ids = model.generate(input_features.to("cuda"), max_new_tokens=128)
pred_text = processor.decode(generated_ids[0], skip_special_tokens=True)

print("Pred text:", pred_text)
print("Environment set up successful?", generated_ids.shape[-1] == 20)

1. Pseudo-Labelling

The python script run_pseudo_labelling.py is a flexible inference script that can be used to generate pseudo-labels under a range of settings, including using both greedy and beam-search. It is also compatible with 🤗 Datasets streaming mode, allowing users to load massive audio datasets with no disk space requirements. For more information on streaming mode, the reader is referred to the blog post: A Complete Guide to Audio Datasets.

As of the latest Distil-Whisper release, distil-large-v3, this pseudo-labelling script also performs the added operation of concatenating (or packing) the audio inputs to 30-seconds. Not only does this lead to a WER improvement when using sequential long-form decoding algorithm, but concatenating audios to 30-seconds also improves the throughput during training, since the amount of zero-padding on the audio inputs is minimised.

The following script demonstrates how to pseudo-label the Hindi split of the Common Voice 16.1 dataset with greedy sampling:

#!/usr/bin/env bash

accelerate launch run_pseudo_labelling.py \
  --model_name_or_path "openai/whisper-large-v3" \
  --dataset_name "mozilla-foundation/common_voice_16_1" \
  --dataset_config_name "hi" \
  --dataset_split_name "train+validation+test" \
  --text_column_name "sentence" \
  --id_column_name "path" \
  --output_dir "./common_voice_16_1_hi_pseudo_labelled" \
  --wandb_project "distil-whisper-labelling" \
  --per_device_eval_batch_size 64 \
  --dtype "bfloat16" \
  --attn_implementation "sdpa" \
  --logging_steps 500 \
  --max_label_length 256 \
  --concatenate_audio \
  --preprocessing_batch_size 500 \
  --preprocessing_num_workers 8 \
  --dataloader_num_workers 8 \
  --report_to "wandb" \
  --language "hi" \
  --task "transcribe" \
  --return_timestamps \
  --streaming False \
  --generation_num_beams 1 \
  --push_to_hub

On an 80 GB A100 GPU, the following script takes approximately 5 minutes to concatenate and pre-process the 20 hours of audio data, and a further 10 minutes to transcribe the pseudo-labels. The pseudo-labelled dataset corresponding to this script is available on the Hugging Face Hub under sanchit-gandhi/common_voice_16_1_hi_pseudo_labelled. The WER of the pre-trained Whisper large-v3 model is 17.2% on the test split. We will compare the performance of our distilled model against this number.

There are two noteworthy arguments that configure the dataset concatenation (or packing) process:

1. concatenate_audio: whether or not to concatenate (or pack) the audios to 30-second chunks. The latest Distil-Whisper model, distil-large-v3, highlights the WER improvements obtained using the sequential long-form decoding algorithm when concatenated audios are used. Concatenating audios to 30-seconds also improves the throughput during training, since the amount of zero-padding on the audio inputs is minimised. Hence, it is highly recommended to set --concatenate_audio=True.
2. preprocessing_batch_size: the batch size to use when concatenating (or packing) the audios. Using a larger batch size results in a greater portion of audio samples being packed to 30-seconds, at the expense of higher memory consumption. If you exceed your system's RAM when performing the concatenation operation, reduce the preprocessing_batch_size by a factor of 2 to 250 or even 125.
3. preprocessing_num_workers: the number of multiprocessing workers to use when concatenating the audios. Using more workers will result in faster pre-processing, at the expense of higher memory consumption. Ensure you do not exceed the maximum number of CPUs on your device.

In addition, the following arguments configure the inference of the Whisper model:

1. language: explicitly setting the language token during inference substantially improves the generation performance of the Whisper model, since the model is forced always to predict in the given language. We recommend you set the language to the language you wish to distil the Whisper model on. The only exception is when distilling an English-only model (i.e. where the model id is appended with an .en, e.g. small.en), the language argument should be set to None, since there is no language token used during training/inference.
2. return_timestamps: whether or not to predict timestamps in the pseudo-labels. Timestamp prediction is required should you want your distilled model to be able to predict timestamps at inference time (e.g. for the original OpenAI long-form transcription algorithm). However, the pseudo-labels are marginally less accurate than not using timestamps. We recommend pseudo-labelling with timestamps to ensure the distilled model is as general as possible.
3. attn_implementation: which attention implementation to use for inference. Set to sdpa for PyTorch SDPA, or flash_attn_2 if your hardware supports Flash Attention 2 and you have the package installed.
4. streaming: whether or not to use Datasets' streaming mode. If enabled, the audio data will be streamed from the Hugging Face Hub with no disk space requirements. However, the user is then responsible for adding the pseudo-labels to the dataset script in a follow-up step (see Using Streaming Mode). If set to False, the audio data will be downloaded and pre-processed offline. At the end of pseudo-labelling, the pseudo-labels will be automatically appended to the original dataset, meaning the dataset is ready to be used for the subsequent training step without any additional steps.
5. generation_num_beams: how many beams to use while decoding. In practice, we found the distilled model to perform comparably when the data was pseudo-labelled with generation_num_beams=1 (greedy) or generation_num_beams>1 (beam). This is likely because the WER filter compensates for the lower quality pseudo-labels obtained using greedy search. However, using generation_num_beams=1 gives substantially faster inference time for the pseudo-labelling step, and so we recommend this configuration.

Should you have your own audio dataset, you can first convert it to Hugging Face Datasets format and push it to the Hugging Face Hub. You can then pseudo-label it using the script above, replacing the --dataset_name with the name of your dataset on the Hub.

Otherwise, you may wish to use an open-source dataset already available on the Hugging Face Hub. We provide a summary of the three most popular multilingual datasets in the table below. For more details, refer to the blog post: A Complete Guide to Audio Datasets.

Dataset	Languages	Domain	Speaking Style	License	Text Column	ID Column
Multilingual LibriSpeech	6	Audiobooks	Narrated	CC-BY-4.0	"text"	"id"
Common Voice 16	120	Wikipedia text & crowd-sourced speech	Narrated	CC0-1.0	"sentence"	"path"
VoxPopuli	15	European Parliament recordings	Spontaneous	CC0	"normalized_text"	"audio_id"

To achieve robustness to different distributions of audio data, it is recommended to train on multiple datasets where possible. For example, the above three datasets all have splits for the German language. Thus, if distilling a Whisper model for German, it would be wise to use a combination of the three datasets during training, in order to cover at least three distinct domains (audiobooks, crowd-sourced speech, parliament recordings). You may wish to use a combination of open-source datasets, or a combination of open-source and individually owned datasets to cover multiple distributions and domains. Moreover, if you were to train on low-resource datasets (<500 hours), you could experiment with language mixing to improve robustness.

2. Initialisation

The script create_student_model.py can be used to initialise a small student model from a large teacher model. When initialising a student model with fewer layers than the teacher model, the student is initialised by copying maximally spaced layers from the teacher, as per the DistilBart recommendations.

First, we need to create a model repository on the Hugging Face Hub. This repository will contain all the required files to reproduce the training run, alongside model weights, training logs and a README.md card. You can either create a model repository directly on the Hugging Face Hub using the link: https://huggingface.co/new. Or, via the CLI, as we'll show here.

Let's pick a name for our distilled model: distil-whisper-large-v3-hi. We can run the following command to create a repository under this name:

huggingface-cli repo create distil-whisper-large-v3-hi

We can now see the model on the Hub, e.g. under https://huggingface.co/sanchit-gandhi/distil-whisper-large-v3-hi

Let's clone the repository so that we can place our training script and model weights inside:

git lfs install
git clone https://huggingface.co/sanchit-gandhi/distil-whisper-large-v3-hi

Be sure to change the repo address to https://huggingface.co/<your-user-name>/<your-repo-name>

We can now copy the relevant training scrips to the repository:

cd distil-whisper-large-v3-hi

cp ../distil-whisper/training/create_student_model.py .
cp ../distil-whisper/training/run_distillation.py .

The following command demonstrates how to initialise a student model from the Whisper large-v3 checkpoint, with all 32 encoder layer and 2 decoder layers. The 2 student decoder layers are copied from teacher layers 1 and 32 respectively, as the maximally spaced layers:

#!/usr/bin/env bash

python create_student_model.py \
  --teacher_checkpoint "openai/whisper-large-v3" \
  --encoder_layers 32 \
  --decoder_layers 2 \
  --save_dir "./distil-large-v3-init"

The initialised model will be saved to the sub-directory distil-large-v3-init in our model repository.

Note: You can leverage language transfer by setting --teacher_checkpoint to "distil-whisper/distil-large-v3", see language transfer for more details.

3. Training

The script run_distillation.py is an end-to-end script for loading multiple datasets, a student model, a teacher model, and performing teacher-student distillation. It uses the loss formulation from the Distil-Whisper paper, which is a weighted sum of the cross-entropy and KL-divergence loss terms.

The following command takes the Common Voice dataset that was pseudo-labelled in the first stage and trains the 2-layer decoder model intialised in the previous step. We pass the local path to the pseudo-labelled Common Voice dataset (../common_voice_16_1_hi_pseudo_labelled), which you can change to the path where your local pseudo-labelled dataset is saved.

In this example, we will combine the train and validation splits to give our training set, and evaluate on the test split only. This is purely to demonstrate how to combine multiple pseudo-labelled datasets for training, rather than recommended advice for defining train/validation splits. We advise that you train on the train splits of your dataset, evaluate and tune hyper-parameters on the validation split, and only test the final checkpoint on the test split. Note how multiple training datasets and splits can be loaded by separating the dataset arguments by + symbols. Thus, the script generalises to any number of training datasets.

#!/usr/bin/env bash

accelerate launch run_distillation.py \
  --model_name_or_path "./distil-large-v3-init" \
  --teacher_model_name_or_path "openai/whisper-large-v3" \
  --train_dataset_name "../common_voice_16_1_hi_pseudo_labelled+../common_voice_16_1_hi_pseudo_labelled" \
  --train_split_name "train+validation" \
  --text_column_name "sentence+sentence" \
  --train_dataset_samples "7+4" \
  --eval_dataset_name "../common_voice_16_1_hi_pseudo_labelled" \
  --eval_split_name "test" \
  --eval_text_column_name "sentence" \
  --eval_steps 1000 \
  --save_steps 1000 \
  --warmup_steps 50 \
  --learning_rate 0.0001 \
  --lr_scheduler_type "constant_with_warmup" \
  --timestamp_probability 0.2 \
  --condition_on_prev_probability 0.2 \
  --language "hi" \
  --task "transcribe" \
  --logging_steps 25 \
  --save_total_limit 1 \
  --max_steps 5000 \
  --wer_threshold 20 \
  --per_device_train_batch_size 32 \
  --per_device_eval_batch_size 32 \
  --dataloader_num_workers 8 \
  --preprocessing_num_workers 8 \
  --ddp_timeout 7200 \
  --dtype "bfloat16" \
  --attn_implementation "sdpa" \
  --output_dir "./" \
  --do_train \
  --do_eval \
  --gradient_checkpointing \
  --overwrite_output_dir \
  --predict_with_generate \
  --freeze_encoder \
  --freeze_embed_positions \
  --streaming False \
  --push_to_hub

The above training script will take approximately 3 hours to complete on an 80 GB A100 GPU and yield a final WER of 76%. While the generations are starting to take form, there is still a 59% WER gap to the teacher model. This is hardly surprising give we only have 15 hours of un-filtered data, and closer to just 1.5 hours with data filtering. As mentioned above, using upwards of 1000 hours of data and training for 10k steps will likely yield more competitive performance. For the Distil-Whisper paper, we trained on 21k hours of audio data for 80k steps. We found that upwards of 13k hours of audio data was required to reach convergence on English ASR (see Section 9.2 of the paper), so the more data you have, the better!

Scaling to multiple GPUs using distributed data parallelism (DDP) is trivial: simply run accelerate config and select the multi-GPU option, specifying the IDs of the GPUs you wish to use. The above script can then be run using DDP with no code changes.

Training logs will be reported to TensorBoard and WandB, provided the relevant packages are available. An example of a saved checkpoint pushed to the Hugging Face Hub can be found here: sanchit-gandhi/distil-whisper-large-v3-hi.

There are a few noteworthy data arguments:

1. train_dataset_samples: defines the number of training samples in each dataset. Used to calculate the sampling probabilities in the dataloader. A good starting point is setting the samples to the number of hours of audio data in each split. A more refined strategy is setting it to the number of training samples in each split, however this might require downloading the dataset offline to compute these statistics.
2. wer_threshold: sets the WER threshold between the normalised pseudo-labels and normalised ground truth labels. Any samples with WER > wer_threshold are discarded from the training data. This is beneficial to avoid training the student model on pseudo-labels where Whisper hallucinated or got the predictions grossly wrong. In our English distillation experiments, we found a WER threshold of 10% provides the optimal trade-off between ensuring high-quality transcriptions, and not filtering unnecessary amounts of training data. For multilingual distillation, the threshold should be set in accordance with the WER achieved by the pre-trained model on the test set.
3. streaming: whether or not to use Datasets' streaming mode. Recommended for large datasets, where the audio data can be streamed from the Hugging Face Hub with no disk space requirements.
4. timestamp_probability: the per-sample probability for retaining timestamp tokens in the labels (should they contain them). Retaining some portion of timestamp tokens in the training data is required to ensure the distilled model can predict timestamps at inference time. In our experiments, we found that training on timestamps with high-probability hurts the distilled model's transcription performance. Thus, we recommend setting this to a value below 0.5. Typically, a value of 0.2 works well, giving good transcription and timestamp performance.
5. condition_on_prev_probability: the per-sample probability for conditioning on previous labels. Conditioning on previous tokens is required to ensure the distilled model can be used with the "sequential" long-form transcription algorithm at inference time. We did not experiment with this parameter, but found values around 0.2 to provide adequate performance. OpenAI pre-trained Whisper on with a 50% probability for conditioning on previous tokens. Thus, you might wish to try higher values.

As well as a few noteworthy model arguments that can be configured to give optimal training performance:

1. freeze_encoder: whether to freeze the entire encoder of the student model during training. Beneficial when the student encoder is copied exactly from the teacher encoder. In this case, the encoder hidden-states from the teacher model are re-used for the student model. Stopping the gradient computation through the encoder and sharing the encoder hidden-states provides a significant memory saving, and can enable up to 2x batch sizes.
2. freeze_embed_positions: whether to freeze the student model's decoder positional embeddings. Using the same embed positions as the teacher model, which is designed to handle context lengths up to 448 tokens, helps the student model retain its input id representation up to the full max input length.
3. dtype: data type (dtype) in which the model computation should be performed. Note that this only controls the dtype of the computations (forward and backward pass), and not the dtype of the parameters or optimiser states.
4. freeze_decoder: whether to freeze the student model's decoder. Note that the input tokens embeddings and language modelling head will remain trainable.

And finally, a few noteworthy training arguments:

1. max_steps: defines the total number of optimisation steps (forward + backward pass) during training. To reach convergence, you should use a dataset of at least 1k hours and train for a minimum of 50k steps.
2. lr_scheduler_stype: defines the learning rate schedule, one of constant_with_warmup or linear. When experimenting with a training set-up or training for very few steps (< 5k), using constant_with_warmup is typically beneficial, since the learning rate remains high over the short training run. When performing long training runs (> 5k), using a linear schedule generally results in superior downstream performance of the distilled model.

TODO:

• Template for model cards

4. Evaluation

There are four types of evaluation performed in Distil-Whisper:

1. Short form: evaluation on audio samples less than 30s in duration. Examples include typical ASR test sets, such as the LibriSpeech validation set.
2. Sequential long form: evaluation on audio samples longer than 30s in duration using the original "sequential" long-form algorithm. Examples include entire TED talks or earnings calls.
3. Chunked long form: evaluation on audio samples longer than 30s in duration using the Transformers "chunked" long-form algorithm.
4. Speculative decoding: evaluation on audio samples less than 30s in duration, where a faster, distilled model is used as the assistant to a slower, teacher model.

All four forms of evaluation are performed using the script run_eval.py. Unlike the pseudo-labelling and training scripts, the evaluation script assumes that only one GPU accelerator is used. We can copy the corresponding evaluation script to the model repository using the following command:

cp ../distil-whisper/training/run_eval.py .

Models are assessed jointly using:

1. The word-error rate (WER) metric: measures the number of substitution, deletion and insertion errors relative to the total number of words. A lower WER indicates a more accurate model.
2. The inverse real-time factor (RTFx) metric: measures the ratio of audio input time : model compute time. A higher RTFx indicates a faster model. Note that this metric is WER-dependent, meaning that it makes sense to compare two models' RTFx only at fixed WER performances. Indeed, deletions could lead to early stopping of token generation, resulting in higher WER and lower RTFx.
3. Token generation speed: This refers to the number of tokens generated per second. As with RTFx, this metric is dependent on the WER since token generation time is not linear. By default, this metric is calculated by averaging the total number of generated tokens : generation time (full forward pass of the model) when evaluating on the given test set. However, using the --precise_tok_generation flag will compute this metric separately for a fixed number of tokens.

In all cases, it is particularly important to evaluate the final model on data that is out-of-distribution (OOD) with the training data. Evaluating on OOD data provides insight as to how well the distilled model is likely to generalise to different audio distributions at inference time. In our example, the Common Voice test set is in-distribution (ID) with our training data, since it is taken from the same distribution as the Common Voice training set. Whereas the FLEURS test set is OOD, since it is not used as part of the training set. See Datasets section for recommendations.

Short Form

The script run_eval.py can be used to evaluate a trained student model over multiple short-form validation sets. The following example demonstrates how to evaluate the student model trained in the previous step on the Common Voice test set (ID) and also the FLEURS test set (OOD). Again, it leverages streaming mode to bypass the need to download the data offline:

#!/usr/bin/env bash

python run_eval.py \
  --model_name_or_path "./" \
  --dataset_name "../common_voice_16_1_hi_pseudo_labelled+google/fleurs" \
  --dataset_config_name "default+hi_in" \
  --dataset_split_name "test+test" \
  --text_column_name "sentence+transcription" \
  --batch_size 16 \
  --dtype "bfloat16" \
  --generation_max_length 256 \
  --language "hi" \
  --attn_implementation "sdpa" \
  --streaming

The student model achieves an average WER of TODO% with an RTFx of TODO for a batch size of 16. We can easily adapt the above script to evaluate the teacher model, simply by switching the model_name_or_path to openai/whisper-large-v3, which achieves an average WER of TODO% with an RTFx of TODO. Therefore, for a batch size of 16, the student model is a factor of TODO times faster than the teacher. The WER gap can be closed by training on more data (at least 1k hours) for more training steps (at least 50k).

Sequential Long Form

The original Whisper paper presents a long-form transcription algorithm that sequentially transcribes 30-second segments of audio and shifts the sliding window according to the timestamps predicted by the model. This style of sequential inference is performed directly using the .generate method in Transformers.

The script run_eval.py can be used to evaluate the trained student model on an arbitrary number of long-form evaluation sets using the sequential algorithm. Since we don't have a long-form validation set for Hindi to hand, in this example we'll evaluate the official Distil-Whisper model distil-large-v3 on the TED-LIUM validation set:

#!/usr/bin/env bash

accelerate launch run_eval.py \
  --model_name_or_path "distil-whisper/distil-large-v3" \
  --dataset_name "distil-whisper/tedlium-long-form" \
  --dataset_config_name "default" \
  --dataset_split_name "validation" \
  --text_column_name "text" \
  --batch_size 16 \
  --dtype "bfloat16" \
  --generation_max_length 256 \
  --language "en" \
  --attn_implementation "sdpa" \
  --streaming

Chunked Long Form

Chunked long form evaluation runs on the premise that a single long audio file can be chunked into smaller segments and inferred in parallel. The resulting transcriptions are then joined at the boundaries to give the final text prediction. A small overlap (or stride) is used between adjacent segments to ensure a continuous transcription across chunks.

This style of chunked inference is performed using the pipeline class, which provides a wrapper around the .generate function for long-form inference.

The script run_eval.py can be used to evaluate the trained student model on an arbitrary number of long-form evaluation sets using the pipeline class. Again, in this example we'll evaluate distil-large-v3 on the TED-LIUM validation set:

#!/usr/bin/env bash

python run_eval.py \
  --model_name_or_path "openai/whisper-large-v3" \
  --dataset_name "distil-whisper/tedlium-long-form" \
  --dataset_config_name "default" \
  --dataset_split_name "validation" \
  --text_column_name "text" \
  --use_pipeline \
  --chunk_length_s 25.0 \
  --language "en" \
  --return_timestamps \
  --dtype "bfloat16" \
  --streaming

The argument chunk_length_s controls the length of the chunked audio samples. It should be set to match the typical length of audio the student model was trained on. If unsure about what value of chunk_length_s is optimal for your case, it is recommended to run a sweep over all possible values. A template script for running a WandB sweep can be found under run_chunk_length_s_sweep.yaml.

Speculative Decoding

Speculative decoding, or assisted generation, relies on the premise that a faster, assistant model can be used to speed-up the generation of a slower, assistant model. Speculative decoding mathematically ensures that exactly the same outputs as Whisper are obtained, while being ~2 times faster. This makes it the perfect drop-in replacement for existing Whisper pipelines, since exactly the same outputs are guaranteed.

Distil-Whisper checkpoints can be designed to be efficient assistant models to Whisper for speculative decoding. More precisely, by freezing the encoder during training, the distilled model can share the same encoder weights as Whisper during inference, since the encoder weights are un-changed. In doing so, only the distilled 2-layer decoder has to be loaded in addition to the original Whisper model, which is approximately an 8% increase to the total parameter count, with up to 2x faster inference for low batch sizes. For more details on speculative decoding, the reader is advised to refer to the following blog post: Speculative Decoding for 2x Faster Whisper Inference.

In the example below, we use our distilled model as an assistant to the large-v3 teacher model during inference:

#!/usr/bin/env bash

python run_eval.py \
  --model_name_or_path "openai/whisper-large-v3" \
  --assistant_model_name_or_path "./" \
  --dataset_name "../common_voice_16_1_hi_pseudo_labelled+google/fleurs" \
  --dataset_config_name "default+hi_in" \
  --dataset_split_name "test+test" \
  --text_column_name "sentence+transcription" \
  --batch_size 16 \
  --dtype "bfloat16" \
  --generation_max_length 256 \
  --language "hi" \
  --attn_implementation "sdpa" \
  --streaming

We see that we achieve a WER of TODO%, the same as what we obtained with the large-v3 model, but with an RTFx of TODO, a factor of TODO faster than using the large-v3 model alone. The RTFx value can be improved by training the student on more data and for more training steps, since this will improve the number of predicted tokens that match the teacher predictions.

Recommendations and guidelines

1. Datasets

As explained, ideally, you should aim for ~1000 hours of audio data for training a distilled model via KD. Moreover, you should evaluate your model on out-of-distribution test sets to assess generalization capacities. With at least 1500 hours of audio data for German, Dutch, French and Spanish, 600 hours for Italian, and 300 hours for Portuguese and Polish (which can be supplemented with your own datasets), a good setup to start with is:

• Training datasets: Common Voice 17 and Multilingual Librispeech. Use the train split for training, and the validation and test splits for in-distribution testing.
• Test datasets: VoxPopuli and Fleurs. Use the validation and test splits for out-of-distribution testing.

2. Student model's decoder

2.1 Number of Decoder Layers

We recommend using a 2-layers decoder (see language transfer below). However, you can adjust the number of decoder layers when initializing the student model to balance between inference speed and accuracy. Experimentation has revealed that the Pareto optimal points are with 2, 3, and 4-layers decoders. For indicative results, after 10,000 training steps and inference on an 80GB Nvidia H100 with a batch size of 16 and 20 tokens generation, compared to Whiper large-v3 baseline:

	rel. token gen. speed	ΔWER(%)
2 layers	$3.66$	$-3.5$
3 layers	$3.35$	$-2.3$
4 layers	$3.11$	$-1.8$

2.2 Language Transfer

If you opt for a 2-layers decoder, consider leveraging language transfer by initializing the student model from the distil-large-v3 English distilled model. For French, this method has shown performance improvements of ΔWER=-1.9% (compared to a 2-layers decoder initialized from Whiper large-v3) after 10,000 training steps.

- --teacher_checkpoint "openai/whisper-large-v3" \
+ --teacher_checkpoint "distil-whisper/distil-large-v3" \

3. Language mixing

If you're working with low-resource languages (<500 hours of audio data), consider mixing your training data with a closely related language (for example, mix French and Spanish) to leverage knowledge transfer between languages. Experiments showed that mixing ~400 hours of French (which resulted in a model with poor generalization capacities) with ~500 hours of Spanish improved the model's out-of-distribution performance on French by ΔWER=-7.5%.

To do this:

1. Run pseudo labeling for each training dataset, setting the --language flag to the language of the respective dataset. In the example of mixing French and Spanish, simply modify the given pseudo labeling command with:

   • pseudo labelling the French dataset

     - --dataset_config_name "hi" \
     - --output_dir "./common_voice_16_1_hi_pseudo_labelled" \
     - --language "hi" \
     + --dataset_config_name "fr" \
     + --output_dir "./common_voice_16_1_fr_pseudo_labelled" \
     + --language "fr" \
     

   • pseudo labelling the Spanish dataset

     - --dataset_config_name "hi" \
     - --output_dir "./common_voice_16_1_hi_pseudo_labelled" \
     - --language "hi" \
     + --dataset_config_name "es" \
     + --output_dir "./common_voice_16_1_es_pseudo_labelled" \
     + --language "es" \
     

2. Conduct training on these pseudo-labeled datasets, using the --language flag set to your targeted language. Note that this flag is only used for evaluation purposes, so you set it to the targeted language. The language token used for forwarding the teacher and student model decoders is the one used and saved in pseudo labels during pseudo-labeling, ensuring it's the correct one for the considered sample. In the example of mixing French and Spanish, simply modify the given training command with:

   - --train_dataset_name "../common_voice_16_1_hi_pseudo_labelled+../common_voice_16_1_hi_pseudo_labelled" \
   - --train_split_name "train+validation" \
   - --eval_dataset_name "../common_voice_16_1_hi_pseudo_labelled" \
   - --eval_split_name "test" \
   + --train_dataset_name "../common_voice_17_0_fr_pseudo_labelled+../common_voice_17_0_es_pseudo_labelled" \
   + --train_split_name "train+train" \
   + --eval_dataset_name "../common_voice_16_1_fr_pseudo_labelled" \
   + --eval_split_name "validation" \
   

Overview of Training Methods

1. Fine-Tuning

For fine-tuning, we take the original Whisper checkpoint and train it on one or more datasets using the standard cross-entropy loss. As such, there is no involvement from the teacher checkpoint during training, and so the fine-tuned model is permitted to overfit to the distribution of the training data we provide. This makes it appealing for "low-resource" languages where the original Whisper model performs poorly, since we can boost the performance of the model on a single language by overfitting to that distribution of data. Note that this means the fine-tuned model is prone to loosing its robustness to different audio distributions, which is the trade-off with improving performance on a specified dataset.

As a rule of thumb, fine-tuning is appropriate for languages where the original Whisper model performs > 20% WER, and we have a relatively small quantity of training data available (< 1000 hours). With fine-tuning, we require as little as 10 hours of training data to significantly boost the performance of the Whisper model. For an in-depth guide to fine-tuning Whisper, the reader is advised to refer to the blog post: Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers.

2. Shrink and Fine-Tune

Shrink and fine-tune (SFT) is a knowledge distillation (KD) technique in which we first shrink the teacher model to a smaller student model by copying maximally spaced layers, and then fine-tune the student model on the cross-entropy loss as described above. Typically, we retain the full encoder from the Whisper model and only shrink the decoder. Retaining the entire encoder helps significantly with maintaining Whisper's robustness to different audio distributions (c.f. Section 9.3 of the Distil-Whisper paper).

We can either train the student model on a dataset of (audio, text) pairs as above. Or, we can use the pre-trained Whisper model to generate pseudo-labels for our audio data, and train on the (audio, pseudo-label) pairs.

Pseudo-labels can be used when either:

1. The original text transcriptions are normalised (lower-cased or no punctuation): the Whisper generated pseudo-labels contain both punctuation and casing, and so can be used as a substitute for the normalised transcriptions
2. The pre-trained Whisper model achieves < 20% WER on the languages: we then know the majority of the pseudo-labels will be accurate enough for us to train on.

They are not recommended when both of the following are true:

1. The original text is punctuated and cased
2. The pre-trained Whisper model achieves > 20% WER on the languages: in this case, we want to overfit to the particular distribution of the language, and so train directly on the original text data

To discard inaccurate pseudo-labels during training, we employ a simple WER heuristic to filter our pseudo-labelled training data. We first normalise the original text and the pseudo-labelled text using the Whisper normaliser. If the WER between the normalised text exceeds a 10% WER threshold, we discard the training sample. Else, we retain it for training. Section 9.1 of the Distil-Whisper paper demonstrates the importance of using this threshold for training.

3. KL Divergence

In the KL Divergence setting, the student model is initialised by shrinking the teacher as before, and then trained to match the predictions of the teacher during training.

Summary of Methods

The following table summarises the two training paradigms: fine-tuning and knowledge distillation (KD). It suggests minimum values for the pre-trained WER / training data to achieve reasonable performance:

Method	Pre-Trained WER / %	Training Data / h
Fine-tuning	> 20	< 1000
KD	< 20	> 1000

Acknowledgements

• OpenAI for the Whisper model and original codebase
• Hugging Face 🤗 Transformers for the Whisper model implementation
• Google's TPU Research Cloud (TRC) program for Cloud TPU v4s used to train the official Distil-Whisper models
• The Hugging Face 🤗 cluster for enabling experimentation with the PyTorch scripts

Citation

If you use this code-base, please consider citing the Distil-Whisper paper:

@misc{gandhi2023distilwhisper,
      title={Distil-Whisper: Robust Knowledge Distillation via Large-Scale Pseudo Labelling}, 
      author={Sanchit Gandhi and Patrick von Platen and Alexander M. Rush},
      year={2023},
      eprint={2311.00430},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}