yuriyvnv/whisper-tiny-high-mixed-pt

Whisper-Tiny Portuguese - High-Quality Filtered Synthetic Data (Best Tiny Configuration)

This model is a fine-tuned version of openai/whisper-tiny for Portuguese automatic speech recognition (ASR). It was trained on Common Voice 17.0 Portuguese combined with WAVe-filtered high-quality synthetic speech data using a strict threshold (q ≥ 0.8).

Purpose

This model represents the best configuration for Whisper-Tiny Portuguese, achieving a 1.39 percentage point improvement over the CV-only baseline. However, the paper emphasizes that this gain is modest:

"The Portuguese Whisper-Tiny model achieves its lowest test WER of 29.33% using the high-quality filtered subset, an improvement of just 1.39 percentage points over the Common Voice baseline of 30.72%. This modest gain offers limited justification for the additional data filtering and preprocessing overhead."

This model demonstrates that while high-quality filtering provides the best results for Tiny, the improvement is marginal compared to the dramatic gains seen with Large-v3 models (+32.6%).

Model Details

Property	Value
Base Model	openai/whisper-tiny
Language	Portuguese (pt)
Task	Automatic Speech Recognition (transcribe)
Parameters	39M
Training Data	Common Voice 17.0 + High-Quality Synthetic (q ≥ 0.8)
Total Training Samples	29,178
Sampling Rate	16kHz

Evaluation Results

This Model (whisper-tiny-high-mixed-pt)

Metric	Value
Validation Loss	0.4481
Validation WER	26.74%
Test WER (Common Voice)	29.33%
Test WER (MLS)	44.18%
Best Checkpoint	Step 350
Max Training Steps	575

Comparison with Other Training Configurations (Whisper-Tiny Portuguese)

Training Data	Max Steps	Val Loss	Val WER	Test WER (CV)	Test WER (MLS)
Common Voice Only	430	0.4463	27.05%	30.72%	45.83%
High-Quality (q ≥ 0.8) + CV	575	0.4481	26.74%	29.33%	44.18%
Mid-High (q ≥ 0.5) + CV	805	0.4550	26.95%	30.11%	47.25%
All Synthetic + CV	860	0.4517	28.06%	29.84%	46.54%

Key Performance Highlights

• Best Tiny configuration: Lowest Test WER (29.33%) and MLS WER (44.18%)
• Modest improvement: Only 1.39% better than baseline on in-domain
• Best cross-domain: 44.18% MLS WER (best among Tiny configurations)
• Quality threshold matters: Strict q ≥ 0.8 filtering provides best results for Tiny

Tiny vs Large: Synthetic Data Impact

The contrast with Large-v3 models illustrates the architectural capacity limitation:

Model	High-Quality Synthetic Impact	Test WER	vs Baseline
Whisper-Tiny	Best config, but marginal	29.33%	+1.39% (4.5% relative)
Whisper-Large-v3	Dramatic improvement	7.94%	+3.84% (32.6% relative)

For Large-v3, high-quality synthetic data reduces WER by 32.6%. For Tiny, the same approach yields only 4.5% relative improvement.

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 ≥ 0.8)	7,312	Strictly WAVe-filtered TTS audio (high quality only)
Total	29,178

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 strict threshold retains only the top 33.3% of synthetic samples, which proves optimal for Tiny models that cannot handle noisier data.

Training Procedure

Hyperparameters

Parameter	Value
Learning Rate	5e-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-tiny-high-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-tiny-high-mixed-pt")
model = WhisperForConditionalGeneration.from_pretrained("yuriyvnv/whisper-tiny-high-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 ideal when:

• Best Tiny accuracy needed: 29.33% WER (best among Tiny configurations)
• Resource-constrained deployment: 39M parameters for edge devices
• Cross-domain robustness for Tiny: Best MLS performance (44.18%)
• Quality-filtered augmentation available: Have WAVe-scored synthetic data

Consider alternatives based on your needs:

• whisper-tiny-cv-only-pt: Simpler setup, only 1.39% worse
• whisper-small-cv-only-pt: Better accuracy (13.87% WER)
• whisper-large-v3-high-mixed-pt: Best overall (7.94% WER)

Research Implications

This model demonstrates an important finding:

High-quality filtering is necessary but not sufficient for smaller models.

For Tiny models:

• Quality filtering (q ≥ 0.8) is the only configuration that helps
• Mid-high quality (q ≥ 0.5) actually hurts performance vs baseline
• Unfiltered data provides worse results than strict filtering
• The improvement is marginal regardless of filtering quality

Recommendation: For resource-constrained deployments, the baseline CV-only model may be more practical given the marginal 1.39% improvement doesn't justify the additional preprocessing complexity.

Limitations

• Lower accuracy than larger models: 29.33% vs 7.94% (Large-v3)
• Marginal improvement over baseline: Only 1.39 percentage points
• Limited capacity: Cannot fully leverage synthetic data benefits
• Domain specificity: Optimized for general Portuguese
• Dialect coverage: Performance may vary across Portuguese regional variants

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-tiny
• 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