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

📍 FOSS Meetup, Kochi @ KeyValue, Smart Kochi.

Kurian Benoy

Saturday, June 24, 2023


  • 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


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


  • 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.

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)

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.

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.

Awesome community plugins

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)

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.

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 🗣

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

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

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

thennal/whisper-medium-ml model card readme

I was not convinced

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

Last commit in thennal/whisper-medium-ml

Objective of my benchmarking

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


  • Common Voice 11 malayalam subset
  • SMC Malayalam Speech Corpus

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)

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.

I wanted to build something new

Time for a new adventure

Libraries I used for benchmarking

  • Dependencies:

    • transformers
    • datasets
    • jiwer
    • whisper_normalizer
    • numerize pandas librosa soundfile
  • Development library:

    • nbdev
    • Jupyter Lab

Loading the dataset for benchmarking

def load_common_voice_malayalam_dataset():
    dataset = load_dataset(
    dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
    dataset =
    dataset = dataset.filter(is_target_text_in_range, input_columns=["norm_text"])
    return dataset

Benchmarking a particular model weight in common voice

    [], [], [], []

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.
    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):
    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)
    print(f"The WER of model: {rwer}")

    rcer = cer(references, predictions)
    rcer = round(100 * rcer, 2)
    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)}")
    df["model_size"] = get_model_size(whisper_asr.model)
    save_name = model_name.split("/")

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

Future Ideas for Benchmarking

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

Open LLM leaderboard in huggingface spaces


  • In Malayalam we have achieved phenomenal results for fine tuned whisper models.
  • The best model after benchmarking is: thennal/whisper-medium-ml
  • I think their seems to be a good ASR model suitable for production use-cases.
  • You can also do it in your own language especially if it is a low resource language.

Thanks to