gilbert-whisper-distil-fr-v0.2

Gilbert-Whisper-Distil-FR-v0.2 — Distilled Model for Production and Research

Overview

Gilbert-Whisper-Distil-FR-v0.2 is a distilled version of Whisper Large V3, optimized for French speech recognition and designed for production deployment and research acceleration within the Gilbert project ecosystem. This model provides 2-4x faster inference while maintaining performance close to the full Large V3 model, making it ideal for real-time applications and cost-effective production systems.

Important Notice on Intellectual Property:

• This distilled model (MEscriva/gilbert-whisper-distil-fr-v0.2) is distributed under the MIT License, allowing research and commercial use.
• All derivative models, fine-tuned variants, and specialized models developed from this distilled model as part of the Gilbert project are the exclusive intellectual property of Lexia France.
• While this model can be used freely under MIT terms, any models built upon it for the Gilbert project are proprietary and subject to separate licensing terms.

---

Research Context

The Gilbert project requires both high-precision baseline models (Gilbert-FR-Source) and production-optimized models for different deployment scenarios. This distilled model serves as the fast inference baseline for:

• Real-time transcription in meeting applications
• Cost-effective batch processing at scale
• Speculative decoding acceleration (2x speedup with identical output)
• Resource-constrained environments (limited GPU memory, edge devices)

This model complements the full-size Gilbert-FR-Source baseline, providing researchers and developers with a speed-optimized alternative for production deployments.

---

Model Details

Architecture

• Base Model: OpenAI Whisper Large V3 (distilled)
• Distillation Method: Patient teacher distillation with 2 decoder layers
• Encoder: Identical to Whisper Large V3 (unchanged, shared with teacher)
• Decoder: 2 layers (reduced from full model)
• Framework: Compatible with Hugging Face Transformers, OpenAI Whisper, Faster Whisper, Whisper.cpp, CTranslate2, ONNX Runtime, and MLX
• Model Size: ~3.0 GB (full precision, but more efficient inference)

Key Characteristics

• Language: French (primary), with multilingual capabilities
• Context Length: Long-form audio support (up to 30 minutes per segment)
• Training: Extended to 30-second audio segments to maintain long-form transcription abilities
• Output: Text transcription with word-level timestamps
• Performance: Optimized for French speech recognition with speed-accuracy trade-off

Distillation Details

• Teacher Model: OpenAI Whisper Large V3
• Training Data: 22,000+ hours of French speech data
• Training Schedule: 160 epochs with aggressive data augmentation
• Method: "Patient teacher" distillation approach
• Segment Length: 30-second segments (preserving same speaker)
• Timestamp Training: 50% of segments trained with timestamps

---

Intended Use

Production Deployment

This model is optimized for:

1. Real-time Applications: Fast inference for live transcription
2. Batch Processing: Cost-effective processing of large audio volumes
3. Resource-Constrained Environments: Lower memory footprint (~2-3 GB GPU)
4. Speculative Decoding: Use as draft model for 2x speedup with identical output

Research and Development

This model serves as:

1. Speed Baseline: Reference point for fast inference research
2. Production Baseline: Starting point for production-optimized fine-tuning
3. Comparative Studies: Benchmark against full-size models (speed vs accuracy)
4. Speculative Decoding Research: Draft model for advanced decoding strategies

Use Cases

• ✅ Meeting Transcription: Fast processing of professional meetings
• ✅ Long-form Audio: Efficient transcription of 30-120 minute sessions
• ✅ Real-time Systems: Live transcription with low latency
• ✅ Cost-Sensitive Applications: Reduce inference costs by 2-4x
• ✅ Edge Deployment: Run on devices with limited GPU memory

---

Performance Benchmarks

Speed Performance

Accuracy Performance

The distilled model maintains performance close to the full model on most tasks:

• Short-form transcription: Competitive with full model
• Long-form transcription: Maintained through 30-second segment training
• French language: Optimized for French speech recognition
• Post-normalization WER: Comparable to full model on standard benchmarks

Note: Exact WER metrics are evaluated on both in-distribution (ID) and out-of-distribution (OOD) datasets. Performance may vary slightly on complex tasks compared to the full model, but the speed-accuracy trade-off is highly favorable for production use.

Speculative Decoding Performance

When used as a draft model with the full Whisper Large V3 as teacher:

• Speed: 2x faster than teacher alone
• Output: Identical to teacher (guaranteed)
• Memory: Only decoder needs to be loaded (encoder shared)
• Use Case: Best of both worlds - speed of distil + accuracy of full model

---

Usage

Installation

pip install transformers torch torchaudio librosa soundfile

Basic Usage with Transformers

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
import torch

model_id = "MEscriva/gilbert-whisper-distil-fr-v0.2"
device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if device == "cuda" else torch.float32

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True
)
model.to(device)

# Process audio
audio_path = "your_audio.wav"
inputs = processor(audio_path, return_tensors="pt", sampling_rate=16000)
inputs = {k: v.to(device) for k, v in inputs.items()}

with torch.no_grad():
    generated_ids = model.generate(
        inputs["input_features"],
        language="fr",
        task="transcribe"
    )

transcription = processor.batch_decode(
    generated_ids,
    skip_special_tokens=True
)[0]

Usage with OpenAI Whisper

import whisper

# Load the distilled model
model = whisper.load_model("large-v3")  # Compatible format

# Transcribe French audio
result = model.transcribe(
    "audio.wav",
    language="fr",
    task="transcribe"
)

print(result["text"])

Speculative Decoding (Advanced)

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
import torch

# Load both models
draft_model_id = "MEscriva/gilbert-whisper-distil-fr-v0.2"
teacher_model_id = "openai/whisper-large-v3"  # or Gilbert-FR-Source

processor = AutoProcessor.from_pretrained(draft_model_id)
draft_model = AutoModelForSpeechSeq2Seq.from_pretrained(draft_model_id)
teacher_model = AutoModelForSpeechSeq2Seq.from_pretrained(teacher_model_id)

# Use speculative decoding for 2x speedup with identical output
# (Implementation details depend on your decoding framework)

---

Comparison with Other Gilbert Models

Gilbert-FR-Source (Full Model)

When to Use Each

Use Distilled Model if:

• ✅ Speed is critical (real-time, batch processing)
• ✅ Resources are limited (GPU memory, costs)
• ✅ Production deployment
• ✅ Speculative decoding desired

Use Full Source Model if:

• ✅ Maximum accuracy required
• ✅ Research baseline needed
• ✅ Complex audio conditions
• ✅ Resources available

---

Research Methodology

Distillation Approach

This model uses patient teacher distillation:

1. Teacher Model: OpenAI Whisper Large V3 (frozen)
2. Student Model: Reduced decoder (2 layers) with same encoder
3. Training: Extended schedule (160 epochs) with aggressive augmentation
4. Data: 22,000+ hours of French speech, 30-second segments
5. Objective: Maintain performance while reducing inference cost

Evaluation Standards

• Speed Metrics: Inference time, throughput, memory usage
• Accuracy Metrics: WER, CER, BLEU (post-normalization)
• Long-form: Evaluation on 30+ minute audio segments
• Comparison: Against full model and other baselines

Versioning

• Current Version: v0.2 (Extended training for long-form)
• Base Version: v0.1 (Initial distillation)
• Future Versions: Production-optimized variants will reference this baseline

---

Limitations

This distilled model inherits limitations from the distillation process:

1. Accuracy Trade-off: Slightly lower accuracy than full model on complex tasks
2. Complex Audio: May struggle more than full model on very challenging audio
3. Out-of-Distribution: Performance may degrade more on OOD data
4. Long-form Edge Cases: Very long segments (>60 min) may show more degradation

However, for most production use cases, the speed-accuracy trade-off is highly favorable.

---

Future Research Directions

Planned Gilbert Distilled Variants

1. Gilbert-Distil-Meetings-v1

   • Fine-tuned on meeting data
   • Optimized for multi-speaker scenarios
   • Production-ready for meeting transcription

2. Gilbert-Distil-Longform-v1

   • Enhanced long-form capabilities
   • Better context stability
   • Optimized for 30-120 minute sessions

3. Gilbert-Distil-Accents-v1

   • Robustness to regional accents
   • Fine-tuned on diverse French accents
   • Production deployment for international use

All future Gilbert models are the exclusive intellectual property of Lexia France and will include detailed evaluation reports.

---

Intellectual Property and Licensing

License for This Model

This distilled model (MEscriva/gilbert-whisper-distil-fr-v0.2) is distributed under the MIT License, allowing:

• ✅ Commercial use
• ✅ Modification
• ✅ Distribution
• ✅ Private use
• ✅ Patent use

See the LICENSE file for full terms.

Intellectual Property Notice

Important: While this model is available under MIT License:

• All derivative models, fine-tuned variants, and specialized models developed as part of the Gilbert project are the exclusive intellectual property of Lexia France.
• Use of this model for Gilbert project development implies acceptance of these IP terms.
• Commercial use of Gilbert project derivatives requires separate licensing agreements.

For licensing inquiries regarding Gilbert project models, contact: [email protected]

---

Citation

If you use this distilled model in your research, please cite:

@software{gilbert_distil_2024,
  title={Gilbert-Whisper-Distil-FR-v0.2: Distilled Model for Production and Research},
  author={MEscriva and Lexia France},
  year={2024},
  url={https://huggingface.co/MEscriva/gilbert-whisper-distil-fr-v0.2},
  version={0.2},
  note={Distilled Whisper model for fast French speech recognition}
}

---

Acknowledgments

This distilled model is based on:

• OpenAI Whisper Large V3 (MIT License)
• bofenghuang/whisper-large-v3-distil-fr-v0.2 (Original distillation work)

We acknowledge the contributions of:

• OpenAI for developing and open-sourcing Whisper
• Bofeng Huang for the French distillation work
• Hugging Face for implementing Whisper in Transformers and creating Distil-Whisper
• The open-source community

---

Contact

For research collaboration, evaluation access, or technical inquiries:

• Website: https://gilbert-assistant.fr
• Email: [email protected]
• Repository: https://huggingface.co/MEscriva/gilbert-whisper-distil-fr-v0.2

---

Changelog

Version 0.2 (2024-12-19)

• Initial Gilbert release
• Based on bofenghuang/whisper-large-v3-distil-fr-v0.2
• Extended training for long-form transcription (30-second segments)
• Patient teacher distillation method
• Multiple format support (Transformers, Whisper, Faster Whisper, etc.)

---

© 2024 Lexia France. All rights reserved for Gilbert project derivatives.