Whisper-Small Portuguese - Common Voice Only (Baseline)

This model is a fine-tuned version of openai/whisper-small for Portuguese automatic speech recognition (ASR). It was trained exclusively on Common Voice 17.0 Portuguese without any synthetic data augmentation, serving as the baseline for evaluating the impact of synthetic speech in ASR training.

Purpose

This baseline model demonstrates the performance achievable with the smaller Whisper-Small architecture (244M parameters) using only real, crowdsourced speech data. It serves as a reference point to evaluate:

• The effectiveness of synthetic data augmentation for smaller model architectures
• The limitations of compact models in leveraging synthetic speech
• Comparison with Large-v3 models to understand scaling effects

Key Finding: Unlike the Large-v3 models which show significant improvements with synthetic data, the Small model shows only marginal benefits from synthetic augmentation, achieving the best overall performance with Common Voice data alone.

Model Details

Evaluation Results

This Model (whisper-small-cv-only-pt)

Comparison with Synthetic Data Augmentation (Whisper-Small Portuguese)

Key Performance Characteristics

• Best in-domain performance: 13.87% Test WER (best among all Small configurations)
• Best validation loss (0.2000) - optimal convergence
• Fastest training: Fewest steps (430) among all configurations
• Demonstrates architectural limitation: Synthetic data provides marginal or negative benefit for smaller models

Why Synthetic Data Doesn't Help Small Models

The paper's findings reveal a fundamental architectural limitation:

"The Tiny and Small variants of Whisper exhibit only marginal benefits from synthetic data augmentation, revealing the limitations imposed by reduced model capacity."

Key Insights:

1. Limited capacity: Compact models (244M params) struggle to disentangle subtle acoustic differences between natural and synthetic speech
2. Conflicting signals: Unlike Large-v3 (1550M params), smaller models become overwhelmed by increased acoustic variability
3. Diminishing returns: Synthetic data introduces complexity that these models cannot effectively reconcile
4. Best strategy: For Small models, focus on high-quality real data rather than synthetic augmentation

Training Data

Dataset Composition

Training Procedure

Hyperparameters

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-cv-only-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-cv-only-pt")
model = WhisperForConditionalGeneration.from_pretrained("yuriyvnv/whisper-small-cv-only-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:

• Resource-constrained deployment: Need smaller model size (244M vs 1550M params)
• Portuguese ASR with limited compute: Best Small model performance without synthetic data overhead
• Baseline comparison: Reference for evaluating synthetic data impact on smaller architectures
• Production efficiency: Fastest inference among Portuguese models

Consider Large-v3 models for better accuracy:

• whisper-large-v3-high-mixed-pt: 7.94% WER (43% better)
• whisper-large-v3-cv-capes-filtered-pt: 6.89% MLS WER (78% better cross-domain)

Small vs Large: Architectural Impact

Limitations

• Lower accuracy than Large models: 13.87% vs 7.94% WER (Large-v3)
• Limited synthetic data benefit: Adding synthetic data doesn't improve performance
• Domain specificity: Optimized for Common Voice-style speech
• 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}
}

@article{perezhohin2026wave,
  title={WAVe: Word-aligned verification of synthetic speech for ASR},
  author={Perezhohin, Yuriy and Castelli, Mauro},
  journal={Information Sciences},
  pages={123591},
  year={2026},
  publisher={Elsevier}
}

References

• Base Model: openai/whisper-small
• Training Data: mozilla-foundation/common_voice_17_0
• 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