OpenAI Whisper and it’s amazing power to do fine-tuning demonstrated on my mother-tongue

Kurian Benoy

Sunday, October 1, 2023

Outline

• What is OpenAI Whisper?
• Features of OpenAI Whisper
• What is Fine-tuning and how to fine-tune Whisper?
• About my mother tongue
• Methodology of benchmarking whisper models
• Results on benchmarking Whisper model
• Future Ideas & Conclusion

$whoami

• AI Engineer & Team Lead @ sentient.io
• Volunteer @ Swathanthra Malayalam Computing(SMC)
• FOSS enthusiast
• Not affiliated to OpenAI

Disclaimer

• This talk is not generated.
• If I use something generated I will explicitly mark as from an LLM.

OpenAI Whisper

• I think Whisper¹ is the most under-rated model released by OpenAI.
• It was open-sourced on September 21, 2022 by releasing the inference code and pre-trained model weights.

1. According to research paper p.2, the name Whisper is an abbrevation for WSPSR: Web-scale Supervised Pretraining for Speech Recognition.

About OpenAI Whisper Model

• Whisper is a computer program which can listen to people talking and write down what they say. (Automatic Speech Recognition Model)
• Whisper can understand people speaking different languages and can even translate what they say into English. (Supports transcription and translation to English)

Note: This was generated with GPT-4 with prompt: explain what is openai whisper to a 5 year old kid?

Whisper Models

Size	Parameters	Required VRAM	Relative speed
tiny	39 M	~1 GB	~32x
base	74 M	~1 GB	~16x
small	244 M	~2 GB	~6x
medium	769 M	~5 GB	~2x
large	1550 M	~10 GB	1x

English Speech Recognition

Whisper is competitive with state of art commercial and open source systems. Diagram from whisper research paper p.9

Multi-lingual Speech recognition

• Whisper model is trained on 99 languages
• OpenAI Whisper API supports just 57 languages as some languages performance are not really good.

• More details can be found in Section 3.4 of Whisper research paper pp.6 - 8.
• Zero-shot Whisper improves performance on Multilingual LibriSpeech(MLS) but is still significantly behind both Maestro, XLS-R, and mSLAM on VoxPopuli. (Whisper research paper p.7)

Runs in almost any device

• Since Whisper followed the open source route, whisper.cpp developed by Georgi Gerganov which is a port of OpenAI’s Whisper model in C/C++.

- It supports the below platforms:

1. Mac OS (Intel and ARM)
2. iOS
3. Android
4. Linux/Free BSD
5. Web Assembly etc.

Runs faster now

• JAX implementation of OpenAI Whisper model upto 70x speedup on TPU using Whisper JAX
• 4x faster with same accuracy using faster-whisper
• Useful transformer makes 2x faster than faster-whisper.

Awesome community plugins

• Word-level time stamps with whisper-timestamped,whisperX etc.
• Fine-Tune Whisper is achieving SOTA in lot of languages
• Speaker diarization
• Audio classification using OpenAI’s Whisper

For more checkout awesome list by Sindre Sorhus

Thanks to Ramsri Goutham article on 7 must know libraries and addons on top of Whisper

What is fine tuning?

Given a pre-trained model, which is a large model which is trained on a very specific task. If we want to fit it into our specific dataset we will train and use the pre-trained model to build a new model which works very well for our task.

Picture from fast.lesson covering steps in finetuning a text classifier model

Fine tuning is still relevant

What are steps for fine-tuning Whisper?

Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers

What are steps for fine-tuning Whisper?

1. Preparing Environment
2. Load dataset
3. Prepare Feature Extractor, Tokenizer and Data
4. Training and evaluation
5. Building a demo(optional)

• Based on article Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers
• More details on fine tuning Whisper

Malayalam performance in whisper paper

Model	WER
tiny	102.7
base	122.9
small	104.8
medium	137.8
large-v1	107.1
large-v2	103.2

Appendix D2.2.2 CommonVoice 9 dataset results (Whisper research paper p.23).

About Malayalam

• Malayalam is my mother tongue.
• Native speakers: 38+ million.(according to 2011 census)
• Spoken in: Kerala, Lakshadweep, Puducherry, wherever Malayalees are living.

Malayalam is morphologically complex language

Comparison of Malayalam TTR with that of European Union Constitution Corpus and DoE-CIIL Corpus from K. Manohar et al.

Picture from Quantitative Analysis of the Morphological Complexity of Malayalam Language by K. Manohar et al. For more information check this paper.

Whisper Event

• HuggingFace Team conducted a whisper fine tuning event for 2 weeks from 5th December 2022 to 19th December 2022. The results were out on 23rd December 2022.
• The goal was to to fine-tune the Whisper model to build state-of-the-art speech recognition systems in the languages of our choice 🗣

Whisper Event huggingface page

Malayalam models produced in Whisper Event

• For the language Malayalam, the results are as follows:

Malayalam models performance in whisper event according to leaderboard

Winning models in Malayalam in Whisper Event

• The winning model for Common voice: thennal/whisper-medium-ml
• The winning model for Fleurs: parambharath/whisper-small-ml

I was not convinced

I was sceptical about the winning models becuase of:

1. Achieving 11% WER in Malayalam is astonishing.
2. In Malayalam there is not even a single yard stick to compare. Most of previous works were done in proprietary datasets and not open-sourced.
3. Malyalam is a morpohologically complex language. So even achieving 30% WER is a big deal.

I was not convinced

4. Didn’t trust the Hugging Face way of evaluating models.

thennal/whisper-medium-ml model card readme

I was not convinced

4. Didn’t trust the Hugging Face way of evaluating models.

Last commit in thennal/whisper-medium-ml

Metrics for evaluating ASR models

• ASR evaulation relies on comparission between ground-truth and ASR output.
• Common metrics for ASR evaluation which are popular and good enough¹ are :

1. Word Error Rate(WER)

2. Character Error Rate(CER)

To learn more about ASR evaluation check this blogpost by AWS

1. According to Rethinking Evaluation in ASR: Are our models robust enough? by Likhomanenko T, Xu, Q., Pratap etc.

I wanted to build something new

• New github project for Malayalam ASR Benchmarking

Time for a new adventure

Objective of my benchmarking

• To test whether 10% WER was possible in available academic datasets.

Datasets

• Common Voice 11 malayalam subset
• SMC Malayalam Speech Corpus

Methodology for benchmarking

1. Create as a python library so further whisper-based transformer models can be benchmark.
2. Calculate WER, CER, model size and time taken to benchmark the model for the listed datasets.
3. Build a reproducible approach, so results of benchmarking is stored as dataset.

Libraries I used

• Dependencies:
  • transformers
  • datasets
  • jiwer
  • whisper_normalizer
  • pandas
  • numerize
  • librosa soundfile

Libraries I used

• Development library:

  • nbdev
  • black
  • Jupyter Lab

Loading the dataset for benchmarking

def load_common_voice_malayalam_dataset():
    dataset = load_dataset(
            "mozilla-foundation/common_voice_11_0",
            "ml",
            split="test"
    )
    dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
    dataset = dataset.map(normalise)
    dataset = dataset.filter(is_target_text_in_range, input_columns=["norm_text"])
    return dataset

Benchmarking a particular model weight in common voice

evaluate_whisper_model_common_voice(
    "openai/whisper-large",
    [], [], [], []
)

Evaluating benchmarking code

def evaluate_whisper_model_common_voice(
        model_name: str, # The model name
        werlist: List[float], # WER List
        cerlist: List[float],# CER list
        modelsizelist: List[str], # model size list
        timelist: List[float], # time(s) list
        bs:int =16, # batch size. Default value is 16.
)->None:
    whisper_asr = pipeline(
            "automatic-speech-recognition", model=model_name, device=0
        )
    dataset = load_common_voice_malayalam_dataset()
    
    predictions = []
    references = []
    start = time.time()
    for out in whisper_asr(data(dataset), batch_size=bs):
        predictions.append(normalizer((out["text"])))
        references.append(normalizer(out["reference"][0]))
    end = time.time()

Calculating WER, CER

    ...
    df = pd.DataFrame({"predictions": predictions, "ground_truth": references})
    df["model_name"] = model_name
    df["wer"] = df.apply(lambda row: wer(normalizer(row["ground_truth"]), normalizer(row["predictions"])), axis=1)
    df["cer"] = df.apply(lambda row: cer(normalizer(row["ground_truth"]), normalizer(row["predictions"])), axis=1)
    df["total_time"] = end-start
    rwer = wer(references, predictions)
    rwer = round(100 * rwer, 2)
    werlist.append(rwer)
    print(f"The WER of model: {rwer}")

    rcer = cer(references, predictions)
    rcer = round(100 * rcer, 2)
    cerlist.append(rcer)
    print(f"The CER of model: {rcer}")

Calculating model_size and storing as a dataset

    ...
    print(f"The model size is: {get_model_size(whisper_asr.model)}")
    modelsizelist.append(get_model_size(whisper_asr.model))
    df["model_size"] = get_model_size(whisper_asr.model)
    
    save_name = model_name.split("/")
    print(save_name)
    df.to_parquet(f"{save_name[0]}_{save_name[1]}_commonvoice.parquet")
    
    clear_gpu_memory()

Benchmarked models

• Started with 6 fine-tuned models in Malayalam and compared it with 6 model versions released by OpenAI.

  1. thennal/whisper-medium-ml
  2. parambharat/whisper-tiny-ml
  3. parambharat/whisper-base-ml
  4. parambharat/whisper-small-ml
  5. anuragshas/whisper-large-v2-ml
  6. DrishtiSharma/whisper-large-v2-malayalam

Results on benechmarking in Common Voice dataset

Output from benchmarking tool

WER in Common Voice dataset

Word Error Rate in Common Voice-9 test split

CER in Common Voice dataset

Character Error Rate in Common Voice-9 test split

Results on benechmarking in Malayalam Speech Corpus dataset

Output from benchmarking tool

WER in Malayalam Speech Corpus

Word Error Rate in MSC

CER in Malayalam Speech Corpus

Character Error rate in MSC

Links to Project

Github project

https://github.com/kurianbenoy/malayalam_asr_benchmarking

Links to Project

Benchmarking results

• Results on SMC Malayalam Speech corpus

  https://huggingface.co/datasets/kurianbenoy/

  malayalam_msc_benchmarking

• Results on Common Voice 11

  https://huggingface.co/datasets/kurianbenoy/

  malayalam_common_voice_benchmarking

Future Ideas for Benchmarking

• Something very similar to OpenLLM Leaderboard with results of latest malayalam speech models.
• Should include results for ASR models based on other architectures like Kaldi, Meta’s MMS, Wav2Vec etc.

Open LLM leaderboard in huggingface spaces

Conclusion

• In Malayalam we have achieved phenomenal results for fine tuned whisper models.
• The best model after benchmarking is: thennal/whisper-medium-ml
• You can also do it in your own language especially if it is a low resource language.

Thanks to

1. OpenAI team - Alec Radford, Jong Wook Kim, Christine McLeavey etc. other authors of Whisper paper
2. Creators of CTranslate2 and faster-whisper - Guillaume Klein
3. HuggingFace team - Sanchit Gandhi, Nicolas Patry, Vaibhav Srivastav etc.
4. Kavya Manohar
5. Santhosh Thottingal
6. Thennal D K

7. AbdulMajedRaja RS
8. Georgi Gerganov
9. Ramsri Goutham
10. Wayde Gilliam
11. Other members in SMC.
12. Jarvis Labs

Tributes

Remembering my first Python teacher