Tagarela - A Portuguese speech dataset from podcasts

1. Problem Statement

Despite Portuguese being one of the most widely spoken languages globally, it remains significantly under-resourced in the field of speech technology compared to English.

• Data Scarcity: Existing Portuguese datasets (e.g., CORAA, NURC-SP, MuPe) are orders of magnitude smaller than English counterparts like GigaSpeech (10k hours) or Emilia (139k hours).
• Quality vs. Quantity Trade-off: Existing spontaneous speech corpora are often too noisy or disfluent for Text-to-Speech (TTS) training, which requires clean, single-speaker audio. Conversely, large-scale raw podcast collections (like "Cem Mil Podcasts") lack the necessary preprocessing (diarization, noise reduction, accurate transcription) to be usable for high-performance Automatic Speech Recognition (ASR) or TTS models.
• Bottleneck: This lack of large-scale, high-quality, curated data hinders the development of robust, natural-sounding speech technologies tailored to Portuguese linguistic nuances.

2. Methodology

The authors introduce TAGARELA, a dataset curated from the "Cem Mil Podcasts" collection (76,000+ hours of raw audio). They developed a comprehensive multi-stage preprocessing pipeline to transform raw podcast audio into a high-quality corpus.

A. Data Source and Scale

• Source: 16,806 episodes from 2,094 shows.
• Total Duration: 8,972 hours of audio.
• Dialects: Brazilian Portuguese (8,130h) and European Portuguese (842h).
• Demographics: 70% male speakers, 30% female speakers.

B. The Preprocessing Pipeline (Figure 1)

The pipeline consists of six key stages:

1. Standardization & Segmentation: Audio is converted to FLAC (16kHz, 16-bit, mono). Long recordings are segmented into 5–20 second clips, prioritizing splits at natural silences.
2. Speaker Diarization: Using the pyannote framework, the system identifies and separates speakers to ensure single-speaker segments, crucial for TTS.
3. Overlapping Speech Detection: A dedicated classifier based on Wav2vec2-XLS-R detects and discards segments where multiple speakers talk simultaneously, ensuring clean utterances.
4. Two-Stage Bootstrap Transcription:
   • Seed Corpus: A 1,000-hour subset is transcribed using a commercial API (ElevenLabs Scribe v12).
   • Model Fine-tuning: This seed data is used to fine-tune Whisper large-v3 for pseudo-labeling the remaining data.
   • Quality Control: A Wav2vec2-XLS-R model is also trained on the seed data. The outputs of Whisper and Wav2vec2 are compared; only samples with high agreement (low Word/Character Error Rates between the two) are retained to filter out hallucinations.
5. Audio Enhancement: A Vocos vocoder, repurposed and trained as a denoiser, removes background noise, hiss, and reverb to improve clarity.
6. Labeling:
   • Speaker ID: Clustering via HDBSCAN on RedimNet B6 embeddings creates ~13,368 distinct speaker labels.
   • Dialect Classification: A fine-tuned wav2vec-base model classifies segments as Brazilian or European Portuguese.

C. Dataset Subsets

• Full Set (8,972h): Contains disfluencies, optimized for robust ASR training.
• Clean Set (2,800h): Filtered for high-quality, single-speaker audio, optimized for TTS training.

3. Key Contributions

1. TAGARELA Dataset: The release of the largest public Portuguese speech corpus to date (8,972 hours), rivaling the scale of English's GigaSpeech.
2. Processing Pipeline: A detailed, reproducible pipeline for transforming raw podcast audio into high-fidelity ASR/TTS data, including novel methods for overlap detection and bootstrap transcription validation.
3. Model Validation: Training and evaluation of state-of-the-art (SOTA) ASR and TTS models exclusively on TAGARELA to demonstrate its efficacy.

4. Experimental Results

A. Automatic Speech Recognition (ASR)

• Models Tested: Whisper Large V3, Wav2Vec Large, Distil-Whisper, Parakeet v3, and Parakeet v2 (fine-tuned).
• Best Performance: Parakeet v2 (Fine-tuned) achieved the best results on the TAGARELA test set.
  • Word Error Rate (WER): 15.18%
  • Character Error Rate (CER): 7.09%
• Comparison: The fine-tuned Parakeet v2 significantly outperformed the pre-trained Whisper Large V3 (20.91% WER) and other baselines, proving the dataset's value for training Portuguese-specific models.

B. Text-to-Speech (TTS)

• Models Tested: Orpheus-TTS and Chatterbox, trained on the 2,800-hour clean subset.
• Metrics: WER/CER (Intelligibility) and Mean Opinion Score (MOS) for naturalness.
• Results:
  • Orpheus-TTS: Superior intelligibility (WER: 0.095 ± 0.100).
  • Chatterbox: Slightly higher naturalness (MOS: 4.176 ± 0.983).
  • Ground Truth: MOS of 4.231.
• Analysis: The results highlight a trade-off between linguistic precision and prosody but confirm that the dataset enables the training of high-quality TTS systems, approaching human-level naturalness.

C. Objective Audio Quality

The dataset was evaluated using STOI (Intelligibility), PESQ (Perceived Quality), and SI-SDR (Signal Fidelity). The denoising pipeline significantly improved these metrics compared to raw audio, validating the effectiveness of the enhancement stage.

5. Significance and Conclusion

• Bridging the Gap: TAGARELA addresses the critical resource gap for Portuguese speech technology, providing a dataset scale previously unavailable for the language.
• State-of-the-Art Potential: The experiments demonstrate that models trained on TAGARELA achieve competitive performance, validating the dataset's quality.
• Future Impact: By providing a robust foundation for both ASR and TTS, TAGARELA is expected to accelerate the development of more accurate and natural speech technologies for millions of Portuguese speakers.
• Open Science: The dataset, along with the preprocessing models (overlap detector, denoiser) and checkpoints, is released publicly to foster community development.

In summary, the paper presents a successful end-to-end solution for converting massive, unstructured podcast data into a structured, high-quality resource, effectively enabling the next generation of Portuguese speech AI.