TAGARELA: A Portuguese Speech Dataset From Podcasts

TAGARELA is a large-scale Portuguese speech dataset built from podcast audio and curated for speech technology research, especially Automatic Speech Recognition (ASR) and Text-to-Speech (TTS).

The dataset contains more than 8,972 hours of Portuguese speech derived from the Cem Mil Podcasts collection. It includes Brazilian Portuguese and European Portuguese speech, processed through a pipeline involving audio standardization, segmentation, speaker diarization, overlap detection, transcription generation, and quality enhancement.

News

• 2026: TAGARELA was published at ICASSP 2026.
• 2026: TAGARELA was released publicly on Hugging Face for Portuguese speech research.

Dataset Details

Dataset Description

TAGARELA was created to reduce the resource gap for Portuguese speech technologies. While English has several large-scale speech corpora, Portuguese remains comparatively under-resourced, especially for large-scale ASR and TTS training.

The dataset provides audio-text pairs suitable for speech recognition and speech generation research. The full curated corpus contains:

• 8,972+ hours of Portuguese speech
• 16,806 podcast episodes
• 2,094 podcast shows
• 13,368 distinct speakers
• 8,130 hours of Brazilian Portuguese (pt-BR)
• 842 hours of European Portuguese (pt-PT)
• 6,368 hours from male speakers
• 2,604 hours from female speakers
• Audio segments with an average duration of approximately 9.30 ± 5.49 seconds
• Transcriptions with an average length of approximately 27.69 ± 17.06 words

TAGARELA includes a full subset intended for robust ASR training and a clean-speech subset of approximately 2,800 hours intended for high-quality TTS and speech generation tasks.

Dataset Sources

• Project page: https://fredso.com.br/TAGARELA
• arXiv: https://arxiv.org/abs/2603.15326
• IEEE Xplore (ICASSP 2026): https://ieeexplore.ieee.org/document/11462137

Motivation

Portuguese is spoken by hundreds of millions of people but remains comparatively under-resourced in speech technology. Large-scale, public, and high-quality speech datasets are essential for training robust ASR and TTS systems. TAGARELA was created to support research and development of Portuguese speech technologies at a scale closer to major English speech corpora.

The dataset is intended to support:

• Portuguese ASR research
• Portuguese TTS research
• Speech generation research
• Robust modeling of podcast speech
• Research involving Brazilian Portuguese and European Portuguese
• Evaluation of speech models trained on large-scale Portuguese audio

Dataset Usage

TAGARELA is distributed through Hugging Face Datasets in Parquet format.

Because the dataset is large, streaming is recommended for most use cases.

Install Dependencies

pip install datasets soundfile librosa

If Hugging Face authentication is required in your environment, use:

huggingface-cli login

Load with Streaming

Streaming avoids downloading the full dataset before use.

from datasets import load_dataset

dataset = load_dataset(
    "freds0/TAGARELA",
    split="train",
    streaming=True
)

sample = next(iter(dataset))

print(sample.keys())
print(sample["sentence"])
print(sample["path"])
print(sample["audio"])

Iterate Over Samples

from datasets import load_dataset

dataset = load_dataset(
    "freds0/TAGARELA",
    split="train",
    streaming=True
)

for i, sample in enumerate(dataset):
    audio = sample["audio"]
    text = sample["sentence"]

    print(i)
    print("Text:", text)
    print("Sampling rate:", audio["sampling_rate"])
    print("Audio array shape:", audio["array"].shape)

    if i == 4:
        break

Load Without Decoding Audio

If you only need file paths and transcriptions, you can avoid decoding audio during iteration.

from datasets import load_dataset, Audio

dataset = load_dataset(
    "freds0/TAGARELA",
    split="train",
    streaming=True
)

dataset = dataset.cast_column("audio", Audio(decode=False))

sample = next(iter(dataset))

print(sample["path"])
print(sample["sentence"])
print(sample["audio"])

Download the Dataset Locally

Downloading the full dataset requires substantial disk space. The current dataset metadata reports a download size of approximately 1.21 TB and a dataset size of approximately 1.76 TB.

from datasets import load_dataset

dataset = load_dataset(
    "freds0/TAGARELA",
    split="train"
)

print(dataset)

Use for ASR Fine-Tuning

A typical ASR fine-tuning pipeline can use the audio field as input and the sentence field as the target transcription.

from datasets import load_dataset

dataset = load_dataset(
    "freds0/TAGARELA",
    split="train",
    streaming=True
)

for sample in dataset:
    audio_array = sample["audio"]["array"]
    sampling_rate = sample["audio"]["sampling_rate"]
    transcription = sample["sentence"]

    # Use audio_array and transcription in your ASR training pipeline.
    break

Use for TTS Training

For TTS training, use the sentence field as text input and the corresponding audio field as the target speech signal.

from datasets import load_dataset

dataset = load_dataset(
    "freds0/TAGARELA",
    split="train",
    streaming=True
)

for sample in dataset:
    text = sample["sentence"]
    audio = sample["audio"]

    # Use text and audio in your TTS training pipeline.
    break

Intended Uses

Direct Uses

TAGARELA can be used for:

• Training Portuguese ASR models
• Fine-tuning multilingual or Portuguese speech recognition models
• Training Portuguese TTS models
• Training speech generation systems
• Studying robust speech modeling from podcast audio
• Evaluating Portuguese speech models
• Research on Brazilian and European Portuguese speech
• Research on large-scale speech corpus construction
• Developing preprocessing, filtering, or quality estimation methods for podcast speech

Out-of-Scope Uses

TAGARELA should not be used for:

• Speaker identification of private individuals
• Speaker verification or biometric recognition
• Surveillance or tracking
• Unauthorized voice cloning
• Impersonation or deceptive synthetic media generation
• Applications that violate privacy, copyright, publicity rights, or platform terms of the original podcast sources
• High-stakes decision-making without additional validation

Users are responsible for ensuring that their use of the dataset complies with applicable laws, ethical guidelines, institutional policies, and the terms associated with the original podcast sources.

Dataset Structure

The dataset contains a single split:

DatasetDict({
    train: Dataset({
        features: ["audio", "path", "sentence"],
        num_rows: 7111196
    })
})

Data Fields

• audio: Audio sample loaded as a Hugging Face Audio feature, sampled at 16 kHz.
• path: Path or identifier for the corresponding audio file.
• sentence: Text transcription associated with the audio segment.

Data Splits

Split	Examples
train	7,111,196

Released Fields and Corpus-Level Metadata

The default Hugging Face release contains the fields audio, path, and sentence.

Statistics about dialect, gender, and speakers describe the corpus-level distribution and may not be available as columns in the default dataset split.

Dataset Creation

Source Data

TAGARELA was derived from the Cem Mil Podcasts collection, a large Portuguese podcast corpus containing more than 76,000 hours of diverse, multi-dialect Portuguese audio. The original collection contains raw podcast audio and automatically generated transcripts.

TAGARELA applies additional processing to make the data more suitable for ASR and TTS model training.

Dataset Subsets

Subset	Main task	Size	Description
Full subset	ASR	8,972 hours	Includes audio containing various types of disfluencies, designed for robust automatic speech recognition training.
Clean-speech subset	TTS	2,800 hours	Curated speech-only subset designed for high-quality text-to-speech and speech generation tasks.

Processing Pipeline

TAGARELA was created through a multi-stage processing pipeline:

1. Audio standardization
   Audio files were converted to FLAC format with 16 kHz sample rate, 16-bit depth, and mono channel.

2. Segmentation
   Long-form recordings were segmented into 5-20 second clips at natural silence points to maintain speech cohesiveness.

3. Speaker diarization
   The pyannote framework was applied to identify and label speech segments for each speaker individually.

4. Overlapping speech detection
   A Wav2Vec2-XLS-R classifier was trained to identify and discard segments with overlapping speech.

5. Transcription generation
   A bootstrap strategy was used with ElevenLabs Scribe for a seed corpus, followed by fine-tuning Whisper large-v3 for pseudo-labeling with quality filtering via Wav2Vec2-XLS-R agreement.

6. Quality enhancement
   A Vocos vocoder was repurposed as a denoiser to remove background noise, hiss, and light reverberation.

Annotations

The main annotation is the sentence transcription. These transcriptions are generated through a bootstrap ASR strategy and should be treated as large-scale ASR-generated labels rather than fully manual annotations.

Benchmark Results

Automatic Speech Recognition

Models trained on TAGARELA and evaluated on the Common Voice 17.0 Portuguese test set:

Model	WER (%) ↓
Canary-1B-Flash	7.8
Distil-Whisper	9.2
Parakeet TDT	12.3

Text-to-Speech

Models trained on the 2,800-hour clean-speech subset:

Model	CER (%) ↓	WER (%) ↓	MOS ↑
Orpheus-TTS	19.32 ± 31.64	26.81 ± 35.57	4.00 ± 0.94
Chatterbox	23.73 ± 26.17	31.50 ± 30.05	4.53 ± 0.25

Biases and Representational Considerations

TAGARELA reflects the characteristics of the podcast data from which it was derived. Users should consider the following representational factors:

• The corpus is strongly weighted toward Brazilian Portuguese, with approximately 91% of the audio in pt-BR and 9% in pt-PT.
• The corpus is imbalanced by speaker gender, with approximately 70% of the audio from male speakers and 30% from female speakers.
• Podcast content may overrepresent people, topics, speaking styles, production practices, and social groups that are more likely to publish podcasts.
• Speech style may include informal speech, spontaneous speech, disfluencies, interruptions, and domain-specific vocabulary.
• ASR-generated labels may introduce transcription biases or systematic errors from the models used in the transcription pipeline.

These factors should be considered when training, evaluating, or deploying models based on TAGARELA.

Limitations

TAGARELA is a large-scale automatically processed speech dataset. Users should consider the following limitations:

• Transcriptions are generated through a large-scale ASR-based transcription pipeline and may contain errors.
• Podcast audio can contain background noise, informal speech, disfluencies, interruptions, and topic variability.
• Some remaining alignment errors may affect TTS training.
• Dialect, gender, speaker, and content distributions may reflect the biases of the source podcast collection.
• The dataset may contain sensitive, personal, copyrighted, or opinionated content from podcast speech.
• The dataset should be used with care when training generative speech models.

Ethical Considerations

TAGARELA is based on podcast audio. Although the dataset is intended for research in speech technologies, users should consider privacy, consent, copyright, and potential misuse risks.

The dataset should not be used to build systems for speaker impersonation, unauthorized voice cloning, surveillance, biometric recognition, or deceptive synthetic media generation. Generated speech systems trained on this dataset should include safeguards against misuse.

Consent and Privacy

The dataset is derived from podcast speech. Speakers may not have provided direct consent for all downstream machine learning uses.

Users should treat voice data as sensitive and avoid attempts to identify, profile, contact, track, clone, or impersonate speakers.

Safety Considerations for Generative Models

When TAGARELA is used for speech generation or TTS research, users should include safeguards against misuse, including unauthorized voice cloning, impersonation, and deceptive synthetic media generation.

Distribution

TAGARELA is distributed through the Hugging Face dataset repository in Parquet format.

Redistribution and derivative releases must comply with the dataset license, Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).

Users creating derivatives, filtered subsets, annotations, or adapted versions should preserve attribution and distribute derivative material under the same license when required by CC BY-NC-SA 4.0.

Maintenance

The dataset is maintained through the TAGARELA Hugging Face repository and project page.

Users should check the project page and repository for updates, corrections, model releases, changes to documentation, or additional usage recommendations.

License

TAGARELA is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

This means users are free to share and adapt the material under the following terms:

• Attribution: Users must give appropriate credit, provide a link to the license, and indicate if changes were made.
• NonCommercial: Users may not use the material for commercial purposes.
• ShareAlike: If users remix, transform, or build upon the material, they must distribute their contributions under the same license.

License details: https://creativecommons.org/licenses/by-nc-sa/4.0/

Citation

If you use TAGARELA, please cite:

@INPROCEEDINGS{11462137,
  author={De Oliveira, Frederico Santos and Gris, Lucas Rafael Stefanel and Ferreira, Alef Iury Siqueira and Da Rosa, Augusto Seben and Filho, Alexandre Costa Ferro and Casanova, Edresson and Shulby, Christopher Dane and Sousa, Rafael Teixeira and Silva, Diogo Fernandes Costa and Da Silva Soares, Anderson and Filho, Arlindo Rodrigues Galvão},
  booktitle={ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={TAGARELA - A Portuguese Speech Dataset from Podcasts}, 
  year={2026},
  volume={},
  number={},
  pages={15517-15521},
  keywords={Low earth orbit satellites;Digital audio broadcasting;Broadcasting;Filtering;Filters;Circuits and systems;Digital audio broadcasting;Internet of Things;Avatars;Communication systems;speech processing;text-to-speech;dataset;automatic-speech-recognition},
  doi={10.1109/ICASSP55912.2026.11462137}
}

Authors

Frederico Santos de Oliveira¹, Lucas Rafael Stefanel Gris², Alef Iury Siqueira Ferreira², Augusto Seben da Rosa³, Alexandre Costa Ferro Filho², Edresson Casanova⁴, Christopher Dane Shulby⁵, Rafael Teixeira Sousa¹, Diogo Fernandes Costa Silva², Anderson da Silva Soares², Arlindo Rodrigues Galvão Filho²

¹ Federal University of Mato Grosso (UFMT)
² Federal University of Goiás (UFG)
³ Paulista State University (UNESP)
⁴ NVIDIA
⁵ Elsa Speak

Acknowledgements

This work has been fully funded by the project Research and Development of Algorithms for Construction of Digital Human Technological Components supported by the Advanced Knowledge Center in Immersive Technologies (AKCIT) in partnership with the Federal University of Goiás (UFG).