README.md

---
language:
  - en
license: mit
library_name: transformers
pipeline_tag: automatic-speech-recognition
base_model: openai/whisper-tiny.en
tags:
  - whisper
  - automatic-speech-recognition
  - speech-recognition
  - encoder-decoder
  - onnx
  - transformers.js
  - webgpu
  - wasm
  - browser
  - rtx-5090
---

# Streaming Speech Recognition with Whisper Encoder-Decoder

Browser-deployable automatic speech recognition system built around the Whisper encoder-decoder architecture, with LoRA/PEFT experimentation, robustness benchmarking, ONNX optimization, and static client-side deployment.

> **Production artifact in this repository:** Whisper Tiny English FP32 ONNX browser model selected from measured evaluation evidence.

## Live Resources

- **Live Demo:** https://huggingface.co/spaces/anmol-unitmole/streaming-speech-recognition-whisper-encoder-decoder
- **GitHub Project:** https://github.com/unit-mole/encoder-decoder-projects/tree/main/03-streaming-speech-recognition-whisper-encoder-decoder
- **Model Repository:** https://huggingface.co/anmol-unitmole/streaming-speech-recognition-whisper-encoder-decoder-model


## Evaluation Overview

![Whisper Model Comparison](images/model_comparison.png)

## Evaluation Dashboard

![Whisper Evaluation Dashboard](images/evaluation_dashboard.png)


## Model Selection

The project evaluated Whisper Tiny, Small, and Medium in both pretrained and LoRA fine-tuned configurations.

The final browser model was selected based on **accuracy, latency, memory footprint, real-time factor, and ONNX parity**, rather than assuming that fine-tuning would always improve performance.

### Test-Set Results

| Model | WER | CER | Avg. Latency |
|---|---:|---:|---:|
| Tiny pretrained | **6.09%** | 2.64% | **0.118 s** |
| Tiny LoRA | 16.76% | 6.64% | 0.262 s |
| Small pretrained | **3.46%** | 1.44% | 0.232 s |
| Small LoRA | 6.26% | 2.17% | 0.619 s |
| Medium pretrained | 4.68% | 2.50% | 0.405 s |
| Medium LoRA | 4.23% | **1.37%** | 1.188 s |

All values above come from the project's held-out **1,452-example test evaluation**.

## Why Whisper Tiny for Browser Deployment?

Whisper Small pretrained achieved the lowest overall WER, but Whisper Tiny pretrained was selected as the **browser champion** because it provided a substantially lighter runtime:

- WER: **6.09%**
- CER: **2.64%**
- Average latency: **0.118 s**
- Real-Time Factor: **0.017**
- Peak measured GPU memory: **~151 MB**

This provides a stronger quality/performance trade-off for browser deployment.

## Fine-Tuning Findings

LoRA/PEFT fine-tuning was performed for Tiny, Small, and Medium Whisper models.

Fine-tuning did **not** improve every model:

- Tiny LoRA regressed relative to Tiny pretrained.
- Small LoRA regressed relative to Small pretrained.
- Medium LoRA produced a measurable improvement.

### Successful Medium LoRA Result

| Metric | Medium pretrained | Medium LoRA |
|---|---:|---:|
| WER | 4.68% | **4.23%** |
| CER | 2.50% | **1.37%** |

The Medium LoRA experiment is retained as the strongest fine-tuning result.

**Important:** the Medium LoRA checkpoint is experimental evidence from the broader project. This Hugging Face repository contains the selected Tiny browser deployment artifact.

## ONNX Optimization

The selected Tiny browser model was exported to ONNX and evaluated in both FP32 and Q8 configurations.

| Runtime | WER | Relative WER Change | Prediction Match |
|---|---:|---:|---:|
| PyTorch | 6.12% | Baseline | — |
| **FP32 ONNX** | **5.90%** | **-3.57%** | **98%** |
| Q8 ONNX | 6.45% | +5.36% | 94% |

The project release criterion allowed no more than **2% relative WER regression**.

Therefore:

- ✅ **FP32 ONNX is the production/default browser artifact**
- ⚠️ Q8 is retained only as an experimental optimization artifact

## Robustness Evaluation

The selected Tiny pretrained model was evaluated on **1,800 robustness examples** derived from 300 source recordings.

| Condition | WER |
|---|---:|
| Clean | 6.02% |
| Clipping | 5.80% |
| Low volume | 5.76% |
| Mild Gaussian noise | 6.83% |
| Medium Gaussian noise | 11.15% |
| Heavy Gaussian noise | 21.14% |

Overall robustness WER: **9.45%**

The largest degradation occurs under heavy additive Gaussian noise.

## Browser Architecture

```text
Microphone / Audio Upload
        ↓
Browser Audio Preprocessing
        ↓
Whisper Feature Extraction
        ↓
FP32 ONNX Encoder
        ↓
Autoregressive Decoder
        ↓
Chunk Aggregation
        ↓
Transcript + Timestamps
```

The application performs inference client-side without requiring a Python inference server.

## Browser Features

The deployed application includes:

- Microphone recording
- Audio file upload
- Bundled sample audio
- Audio preview
- Chunked transcription
- Transcription progress
- Final transcript
- Timestamps
- Confidence/result presentation
- Evaluation dashboard
- WebGPU-capable execution
- WASM fallback

## Experimental Environment

Training and evaluation were executed locally using:

- **GPU:** NVIDIA RTX 5090
- **VRAM:** 32 GB
- **CUDA:** 13.0
- **Python:** 3.12
- **PyTorch:** 2.13
- **Transformers:** 4.57.6
- **PEFT:** 0.20.0
- **Mixed Precision:** BF16

## Evaluation Pipeline

The evaluation workflow measures:

- Word Error Rate (WER)
- Character Error Rate (CER)
- Average latency
- P95 latency
- Real-Time Factor
- GPU memory
- Robustness slices
- Noise degradation
- ONNX parity
- Quantization regression
- Browser deployment suitability

## Intended Use

This model and project are suitable for:

- Speech recognition experimentation
- Browser-side ASR demonstrations
- Whisper encoder-decoder research
- ONNX inference experimentation
- WebGPU/WASM deployment experiments
- Machine learning portfolio demonstrations
- Educational use

## Limitations

- The deployment model is English-only.
- The main benchmark is based on LibriSpeech-style English speech.
- Accent coverage in the current benchmark is limited.
- Performance degrades under severe additive environmental noise.
- Browser latency depends on hardware and browser runtime.
- Q8 quantization exceeded the project's quality-regression threshold.
- Speech recognition output may contain transcription errors.

## Responsible Use

This project is intended for experimentation, education, research, portfolio demonstration, and general transcription.

Speech recognition output should **not** be treated as infallible.

Human review should be used before relying on transcripts in:

- Medical applications
- Legal applications
- Emergency situations
- Safety-critical systems
- Security-sensitive workflows
- Other high-stakes environments

## Reproducibility

The complete source code includes:

- Dataset preparation
- Dataset auditing
- LoRA hyperparameter sweeps
- Tiny/Small/Medium Whisper training
- Pretrained vs fine-tuned evaluation
- Robustness benchmarking
- Failure analysis
- ONNX export
- Q8 quantization
- ONNX parity validation
- Browser model selection
- Static web deployment
- Automated Python tests
- JavaScript/browser testing

The full implementation is available here:

https://github.com/unit-mole/encoder-decoder-projects/tree/main/03-streaming-speech-recognition-whisper-encoder-decoder

## License

MIT