cmaree/Bagus-whisper-small-id-onnx

Bagus Whisper Small Indonesian (ONNX)

This is an ONNX-converted version of Bagus/whisper-small-id-cv17, optimized for browser-based inference with Transformers.js.

Model Description

Bagus/whisper-small-id-cv17 is a fine-tuned Whisper Small model for Indonesian automatic speech recognition (ASR). It achieves 5.9% Word Error Rate (WER) on Indonesian speech, making it one of the most accurate Indonesian ASR models available.

This ONNX version provides:

• ✅ Browser-compatible format for client-side inference
• ✅ Multiple quantization levels (q4f16, fp16, int8, q8)
• ✅ Optimized for Transformers.js
• ✅ No server required - runs entirely in the browser

Model Details

• Base Model: OpenAI Whisper Small (242M parameters)
• Language: Indonesian (id)
• Training Data: Mozilla Common Voice 17.0 (Indonesian)
• Performance: 5.9% WER
• License: Apache 2.0
• Original Model: Bagus/whisper-small-id-cv17

Quantization Options

This repository includes multiple quantization levels optimized for different use cases:

Quantization	Encoder	Decoder	Total Size	Use Case
q4f16	51 MB	138 MB	~191 MB	Recommended for browsers - Best balance
fp16	168 MB	293 MB	~463 MB	Higher quality, larger size
int8/q8	87 MB	299 MB	~386 MB	Good quality, moderate size
fp32	337 MB	587 MB	~924 MB	Full precision (reference)

Usage

With Transformers.js (Browser)

import { pipeline } from '@xenova/transformers';

// Create ASR pipeline
const transcriber = await pipeline(
  'automatic-speech-recognition',
  'cmaree/Bagus-whisper-small-id-onnx',
  {
    dtype: 'q4f16',  // Recommended for browser
    device: 'wasm',  // or 'webgpu' if available
  }
);

// Transcribe audio
const result = await transcriber(audioData, {
  language: 'indonesian',
  task: 'transcribe',
});

console.log(result.text);

Custom Wrapper (Recommended)

import {
  AutoProcessor,
  AutoModelForSpeechSeq2Seq,
  AutoTokenizer
} from '@xenova/transformers';

class IndonesianWhisper {
  constructor() {
    this.model = null;
    this.processor = null;
    this.tokenizer = null;
  }

  async load(options = {}) {
    const modelId = 'cmaree/Bagus-whisper-small-id-onnx';
    const device = options.device || 'wasm';

    this.processor = await AutoProcessor.from_pretrained(modelId);
    this.tokenizer = await AutoTokenizer.from_pretrained(modelId);
    this.model = await AutoModelForSpeechSeq2Seq.from_pretrained(modelId, {
      device: device,
      dtype: device === 'webgpu' ? 'fp32' : 'q4f16'
    });
  }

  async transcribe(audioData, options = {}) {
    const processed = await this.processor(audioData, {
      sampling_rate: options.sampling_rate || 16000
    });

    const outputs = await this.model.generate(processed.input_features, {
      max_new_tokens: 128,
      num_beams: 1,
      language: 'indonesian',
      task: 'transcribe',
      return_timestamps: false,
    });

    const transcription = this.tokenizer.batch_decode(outputs, {
      skip_special_tokens: true
    });

    return { text: transcription[0] };
  }
}

// Usage
const whisper = new IndonesianWhisper();
await whisper.load({ device: 'wasm' });
const result = await whisper.transcribe(audioBuffer);
console.log(result.text);

Performance Benchmarks

Accuracy

• Word Error Rate (WER): 5.9%
• Dataset: Common Voice 17.0 (Indonesian test set)
• Comparison:
  • Whisper Tiny (generic): ~25-30% WER
  • Whisper Small (generic): ~10-15% WER
  • Bagus Whisper Small (Indonesian): 5.9% WER ⭐

Inference Speed (Browser)

Tested on typical hardware with WASM backend:

Device	Browser	Speed (RTF)	Notes
Desktop (i7)	Chrome	~0.3x	3-4 seconds for 10s audio
Laptop (M1)	Safari	~0.2x	2-3 seconds for 10s audio
Mobile (Flagship)	Chrome	~0.5x	5-6 seconds for 10s audio

RTF = Real-time Factor (lower is faster)

With WebGPU (when available):

• 2-3x faster inference
• ~0.1-0.15x RTF on modern GPUs

Training Details

Original Model Training

• Base Model: OpenAI Whisper Small
• Training Data: Mozilla Common Voice 17.0 (Indonesian subset)
• Training Framework: PyTorch 2.3.0
• Learning Rate: 1e-05
• Batch Size: 32 (with gradient accumulation)
• Training Steps: 20,000
• Mixed Precision: Native AMP

ONNX Conversion

• Conversion Tool: Transformers.js conversion script
• ONNX Opset: 14
• Quantization: Dynamic quantization (q4f16, int8, fp16)
• Optimization: ONNXSlim for size reduction

Intended Use

Primary Use Cases

• ✅ Real-time Indonesian speech transcription in web browsers
• ✅ Offline voice-to-text applications
• ✅ Voice assistants and chatbots
• ✅ Transcription services
• ✅ Accessibility tools

Out of Scope

• ❌ Speaker diarization (identifying who is speaking)
• ❌ Real-time translation (use with translation models)
• ❌ Audio quality enhancement
• ❌ Languages other than Indonesian

Limitations

1. Audio Quality: Performance degrades with:

   • Background noise
   • Low-quality recordings
   • Multiple speakers
   • Strong accents or dialects

2. Domain Specificity: Trained on Common Voice dataset

   • May perform better on clear, scripted speech
   • May struggle with spontaneous conversations
   • Domain-specific jargon may not be recognized

3. Resource Requirements:

   • Requires ~191MB download (q4f16)
   • Needs modern browser with WASM support
   • May be slow on low-end mobile devices

Ethical Considerations

Bias

• Trained on Common Voice data which may not represent all Indonesian speakers equally
• May have biases toward certain accents, dialects, or demographic groups
• Users should evaluate performance on their specific use case

Privacy

• Model runs entirely in the browser (no data sent to servers)
• Audio processing happens locally
• Users maintain full control over their data

Responsible Use

• Should not be used for surveillance without consent
• Should not be sole basis for high-stakes decisions
• Users should verify transcriptions for critical applications

Citation

Original Model

@misc{bagus2024whisper,
  author = {Bagus},
  title = {Whisper Small Indonesian CV17},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Bagus/whisper-small-id-cv17}
}

Whisper

@misc{radford2022whisper,
  title={Robust Speech Recognition via Large-Scale Weak Supervision},
  author={Alec Radford and Jong Wook Kim and Tao Xu and Greg Brockman and Christine McLeavey and Ilya Sutskever},
  year={2022},
  eprint={2212.04356},
  archivePrefix={arXiv},
  primaryClass={eess.AS}
}

Transformers.js

@software{transformersjs,
  author = {Xenova},
  title = {Transformers.js: State-of-the-art Machine Learning for the web},
  year = {2024},
  url = {https://github.com/xenova/transformers.js}
}

Acknowledgments

• Original Model: Bagus for fine-tuning Whisper on Indonesian
• Base Model: OpenAI for the Whisper architecture
• Dataset: Mozilla Foundation for Common Voice
• Conversion Tools: Xenova for Transformers.js and ONNX conversion scripts
• ONNX Optimization: Microsoft for ONNX Runtime

Contact

For issues or questions about this ONNX version:

• Open an issue on the model repository
• Check Transformers.js documentation

For the original model:

• Visit Bagus/whisper-small-id-cv17

License

This model is released under the Apache 2.0 license, consistent with the original Whisper model and the fine-tuned version.