mjwong/whisper-small-singlish

Whisper small-singlish

Whisper small-singlish is a fine-tuned automatic speech recognition (ASR) model optimized for Singlish. Built on OpenAI's Whisper model, it has been adapted using Singlish-specific data to accurately capture the unique phonetic and lexical nuances of Singlish speech.

Model Details

• Developed by: Ming Jie Wong
• Base Model: openai/whisper-small
• Model Type: Encoder-decoder
• Metrics: Word Error Rate (WER)
• Languages Supported: English (with a focus on Singlish)
• License: Apache-2.0

Description

Whisper small-singlish is developed using an internal dataset of 66.9k audio-transcript pairs. The dataset is derived exclusively from the Part 3 Same Room Environment Close-talk Mic recordings of IMDA's NSC Corpus.

The original Part 3 of the National Speech Corpus comprises approximately 1,000 hours of conversational speech from around 1,000 local English speakers, recorded in pairs. These conversations cover everyday topics and include interactive game-based dialogues. Recordings were conducted in two environments:

• Same Room, where speakers shared a room and were recorded using a close-talk mic and a boundary mic.
• Separate Room, where each speaker was recorded individually using a standing mic and a telephone (IVR).

Audio segments for the internal dataset were extracted using these criteria:

• Minimum Word Count: 10 words

  This threshold was chosen to ensure that each audio segment contains sufficient linguistic context for the model to better understand instructions in Singlish. Shorter segments may bias the model towards specific utterances or phrases, limiting its overall comprehension.

• Maximum Duration: 20 seconds

  This threshold was chosen to provide enough context for accurate transcription while minimizing noise and computational complexity for longer audio segments.

• Sampling Rate: All audio segments are down-sampled to 16kHz.

Full experiments details will be added soon.

Fine-Tuning Details

We applied fine-tuning on a single A100-80GB GPU.

Training Hyperparameters

The following hyperparameters are used:

• batch_size: 64
• gradient_accumulation_steps: 1
• learning_rate: 1e-6
• warmup_steps: 300
• max_steps: 5000
• fp16: true
• eval_batch_size: 16
• eval_step: 300
• max_grad_norm: 1.0
• generation_max_length: 225

Training Results

The table below summarizes the model’s progress across various training steps, showing the training loss, evaluation loss, and Word Error Rate (WER).

Steps	Train Loss	Eval Loss	WER
300	1.4347	0.6711	30.840211
600	0.6508	0.5130	22.538497
900	0.4950	0.3556	18.816530
1200	0.3862	0.3452	17.253038
1500	0.3859	0.3391	17.947677
1800	0.4018	0.3345	16.759187
2100	0.3887	0.3314	16.242452
2400	0.3730	0.3292	15.687331
2700	0.3628	0.3277	15.857115
3000	0.3439	0.3230	15.750816
3300	0.3806	0.3247	15.223008
3600	0.3495	0.3239	15.361788
3900	0.3424	0.3233	15.544122
4200	0.3583	0.3223	15.279849
4500	0.3409	0.3222	15.590628
4800	0.3431	0.3220	15.286493

The final checkpoint is taken from the model trained at 4800 steps.

Benchmark Performance

We evaluated Whisper small-singlish on SASRBench-v1, a benchmark dataset for evaluating ASR performance on Singlish:

Model	WER
openai/whisper-small	148.55%
openai/whisper-large-v3	108.34%
jensenlwt/fine-tuned-122k-whisper-small	92.49%
openai/whisper-large-v3-turbo	56.67%
mjwong/whisper-small-singlish	18.50%
mjwong/whisper-large-v3-turbo-singlish	13.35%

Disclaimer

While this model has been fine-tuned to better recognize Singlish, users may experience inaccuracies, biases, or unexpected outputs, particularly in challenging audio conditions or with speakers using non-standard variations. Use of this model is at your own risk; the developers and distributors are not liable for any consequences arising from its use. Please validate results before deploying in any sensitive or production environment.

How to use the model

The model can be loaded with the automatic-speech-recognition pipeline like so:

from transformers import pipeline
model = "mjwong/whisper-small-singlish"
pipe = pipeline("automatic-speech-recognition", model)

You can then use this pipeline to transcribe audios of arbitrary length.

from datasets import load_dataset
dataset = load_dataset("mjwong/SASRBench-v1", split="test")
sample = dataset[0]["audio"]

result = pipe(sample)
print(result["text"])

Contact

For more information, please reach out to mingjwong@hotmail.com.

Acknowledgements

1. https://www.jensenlwt.com/blog/singlish-whisper-finetuning-asr-for-singapore-unique-english
2. https://github.com/huggingface/community-events/blob/main/whisper-fine-tuning-event/README.md
3. https://medium.com/htx-dsai/finetuning-whisper-for-the-singaporean-home-team-context-a3ae1a6ae809