yuriyvnv/whisper-small-mixed-pt

Whisper-Small Portuguese - Mid-High Quality Filtered Synthetic Data

This model is a fine-tuned version of openai/whisper-small for Portuguese automatic speech recognition (ASR). It was trained on Common Voice 17.0 Portuguese combined with WAVe-filtered synthetic speech data using a balanced quality threshold (q ≥ 0.5), including both high-quality and medium-quality samples.

Purpose

This model tests whether increasing synthetic data volume with moderate quality filtering can benefit smaller model architectures. The results continue to support the finding that Small models cannot effectively leverage synthetic data:

Key Finding: Even with the balanced threshold (q ≥ 0.5) that works best for Large-v3 models, the Small model shows no improvement over the CV-only baseline, reinforcing that model capacity is the fundamental limiting factor.

Metric	CV-Only Baseline	This Model (Mid-High)	Large-v3 (Same Threshold)
Test WER (CV)	13.87%	14.08% (-1.5%)	8.33% (+29.3%)
Test WER (MLS)	30.69%	30.54% (+0.5%)	10.27% (+32.9%)

The contrast with Large-v3 is stark: the same data configuration that produces dramatic improvements for large models provides negligible benefit for small models.

Model Details

Property	Value
Base Model	openai/whisper-small
Language	Portuguese (pt)
Task	Automatic Speech Recognition (transcribe)
Parameters	244M
Training Data	Common Voice 17.0 + Mid-High Quality Synthetic (q ≥ 0.5)
Total Training Samples	41,047
Sampling Rate	16kHz

Evaluation Results

This Model (whisper-small-mixed-pt)

Metric	Value
Validation Loss	0.2100
Validation WER	12.97%
Test WER (Common Voice)	14.08%
Test WER (MLS)	30.54%
Best Checkpoint	Step 300
Max Training Steps	805

Comparison with Other Training Configurations (Whisper-Small Portuguese)

Training Data	Max Steps	Val Loss	Val WER	Test WER (CV)	Test WER (MLS)
Common Voice Only	430	0.2000	12.68%	13.87%	30.69%
High-Quality (q ≥ 0.8) + CV	575	0.2100	12.98%	14.28%	30.40%
Mid-High (q ≥ 0.5) + CV	805	0.2100	12.97%	14.08%	30.54%
All Synthetic + CV	860	0.2100	12.94%	14.22%	30.85%

Key Performance Characteristics

• Best in-domain among augmented: 14.08% Test WER (best among synthetic-augmented Small models)
• Still worse than baseline: 1.5% worse than CV-only (13.87%)
• Marginal cross-domain: 30.54% MLS (only 0.5% better than baseline)
• Largest synthetic dataset: 19,181 synthetic samples used

Small vs Large: Same Data, Different Results

The mid-high quality threshold (q ≥ 0.5) is optimal for Large-v3 but ineffective for Small:

Model	Same Configuration	Test WER (CV)	Test WER (MLS)	Benefit vs Baseline
Whisper-Small	Mid-High + CV	14.08%	30.54%	-1.5% / +0.5%
Whisper-Large-v3	Mid-High + CV	8.33%	10.27%	+29.3% / +32.9%

This ~30 percentage point difference in effectiveness demonstrates that synthetic data benefits are architecture-dependent.

Training Data

Dataset Composition

Source	Samples	Description
Common Voice 17.0 Portuguese	21,866	Real speech from Mozilla's crowdsourced dataset
Synthetic Transcript PT (q ≥ 0.5)	19,181	WAVe-filtered TTS audio (high + medium quality)
Total	41,047

WAVe Quality Distribution (Portuguese Synthetic Data)

Quality Level	Samples	Percentage	Used in This Model
High (q ≥ 0.8)	7,312	33.3%	✓
Medium (0.5 ≤ q < 0.8)	11,869	54.0%	✓
Low (q < 0.5)	2,787	12.7%	✗

This threshold retains 87.3% of the synthetic dataset, the same configuration that achieves best cross-domain performance for Large-v3.

Training Procedure

Hyperparameters

Parameter	Value
Learning Rate	1e-5
Batch Size (Global)	256
Warmup Steps	200
Max Epochs	5
Precision	BF16
Optimizer	AdamW (fused)
Eval Steps	50
Metric for Best Model	eval_loss

Training Infrastructure

• GPU: NVIDIA H200 (140GB VRAM)
• Operating System: Ubuntu 22.04
• Framework: Hugging Face Transformers

Usage

Transcription Pipeline

from transformers import pipeline

transcriber = pipeline(
    "automatic-speech-recognition",
    model="yuriyvnv/whisper-small-mixed-pt",
    device="cuda"
)

result = transcriber("path/to/portuguese_audio.wav")
print(result["text"])

Direct Model Usage

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import librosa

processor = WhisperProcessor.from_pretrained("yuriyvnv/whisper-small-mixed-pt")
model = WhisperForConditionalGeneration.from_pretrained("yuriyvnv/whisper-small-mixed-pt")
model.to("cuda")

audio, sr = librosa.load("path/to/portuguese_audio.wav", sr=16000)
input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features.to("cuda")

predicted_ids = model.generate(input_features)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)

Specifying Language

model.generation_config.language = "pt"
model.generation_config.task = "transcribe"

When to Use This Model

This model is primarily useful for:

• Research purposes: Understanding how model capacity affects synthetic data utilization
• Best among augmented Small: If synthetic data must be used, this is the best configuration
• Comparing architectures: Demonstrating the Small/Large effectiveness gap

For production use, consider:

• whisper-small-cv-only-pt: Best Small model (13.87% WER)
• whisper-large-v3-mixed-pt: Best cross-domain (10.27% MLS)

Research Implications

This model reinforces key findings:

1. Model capacity is fundamental: The same high-quality synthetic data that dramatically improves Large-v3 provides no benefit for Small
2. Diminishing returns don't apply: It's not that Small models benefit less—they don't benefit at all
3. Architecture selection matters: Choose model size first, then decide on synthetic augmentation

Recommendation: For resource-constrained deployments, invest in model optimization (quantization, distillation) rather than synthetic data augmentation.

Limitations

• Lower accuracy than baseline: 14.08% vs 13.87% (worse than CV-only)
• Wasted compute: 87% more training steps for no improvement
• Architecture limitation: Cannot leverage synthetic data effectively
• Domain specificity: Optimized for general Portuguese

Citation

This model is part of research on WAVe (Word-Aligned Verification) for synthetic speech quality assessment. While the WAVe methodology paper is currently under review, please cite our previous work that motivated this research:

@article{perezhohin2024enhancing,
  title={Enhancing Automatic Speech Recognition: Effects of Semantic Audio Filtering on Models Performance},
  author={Perezhohin, Yuriy and Santos, Tiago and Costa, Victor and Peres, Fernando and Castelli, Mauro},
  journal={IEEE Access},
  year={2024},
  publisher={IEEE}
}

References

• Base Model: openai/whisper-small
• Training Data (Real): mozilla-foundation/common_voice_17_0
• Training Data (Synthetic): yuriyvnv/synthetic_transcript_pt
• Whisper Paper: Robust Speech Recognition via Large-Scale Weak Supervision
• Motivating Research: Enhancing ASR with Semantic Audio Filtering (IEEE Access 2024)

License

Apache 2.0