TBOGamer22/BrahuiSpeech-70H-V2

BrahuiSpeech-70H V2

BrahuiSpeech-70H V2 is an approximately 70-hour automatic speech recognition dataset containing 15,626 audio-transcription pairs and 69 hours, 32 minutes, 12 seconds of real-world Brahui (Brahvi) speech.

Brahui (brh) is a low-resource Dravidian language spoken primarily in Balochistan, Pakistan. The dataset covers naturally occurring speech across varied speakers, speaking styles, media domains, and acoustic conditions. Transcriptions use the Perso-Arabic writing system commonly used for Brahui in Pakistan.

V2 improves transcription quality through a second, audio-grounded adjudication pass over high-disagreement samples and removes one confirmed no-speech clip.

V2 Improvements

• —2,507 high-disagreement clips were selected for a second audio review.
• —2,506 clips received usable second-pass adjudications.
• —1,858 transcriptions were corrected during that pass.
• —648 transcriptions were listened to and confirmed unchanged.
• —1 clip containing only laughter and sound effects was removed.
• —The public metadata contains no model predictions, API notes, or private filesystem paths.

The high-disagreement set was identified using a fine-tuned Whisper model and character error rate against the prior transcript. During adjudication, reviewers were given the audio and both transcript hypotheses, with the audio designated as the sole source of truth. Because model predictions were visible during this targeted correction process, any post-adjudication CER against those labels is a quality-control diagnostic, not an unbiased benchmark result.

Dataset Summary

The 70H name denotes the rounded duration class. The exact retained duration is reported above.

Loading

from datasets import load_dataset

dataset = load_dataset("TBOGamer22/BrahuiSpeech-70H-V2")
print(dataset["train"][0])

Each example contains:

audio               Audio
audio_file_name     string
utterance_id        string
text                string
transcription       string
duration_seconds    float
review_stage        string

text and transcription contain identical Brahui labels for compatibility with different ASR training pipelines. review_stage is either audio_reviewed or audio_adjudicated.

Collection Domains

The corpus contains real-world speech gathered from varied media and speaking contexts, including conversations, interviews, discussions, stories, educational material, entertainment, podcasts, news and broadcast-style speech, informal speech, narrative speech, and other online spoken media.

This breadth introduces realistic variation in pronunciation, speaking rate, speaker characteristics, recording devices, background noise, microphone quality, and domain vocabulary.

Transcription And Quality Control

The transcription workflow combined native-speaker transcription, automation, manual transcription and review, and AI-assisted quality assurance. Preparation included:

• —transcript cleanup and Unicode/whitespace normalization;
• —audio validation and file matching;
• —malformed and missing sample removal;
• —Silero VAD speech-presence auditing and complete-silence filtering;
• —full audio-transcript review;
• —targeted second-pass adjudication of high-disagreement samples;
• —removal of a confirmed no-speech sample.

AI systems were used as quality-assurance tools rather than as a substitute for the audio. For second-pass adjudication, instructions explicitly required listening to the full clip, treating both transcript candidates as untrusted, removing unsupported or trailing text, preserving genuine repetitions, and avoiding translation.

Intended Uses

The dataset is intended for automatic speech recognition training and fine-tuning, low-resource speech research, multilingual transfer learning, speech representation learning, and Brahui language-technology development.

This release provides one training split. Researchers should construct leakage-aware validation and test sets appropriate to their source, speaker, and domain assumptions. Do not use the targeted post-adjudication CER as a test-set benchmark.

Limitations

• —Real-world recording conditions are not uniform.
• —Some samples may include background noise, music, reverberation, code-switching, regional pronunciation, or overlapping speech.
• —The corpus does not guarantee balanced representation across dialects, regions, demographics, speakers, or source domains.
• —Automated and AI-assisted review can still leave transcription errors.
• —review_stage describes processing depth, not a numerical quality guarantee.
• —Source-media provenance may create legal, privacy, publicity, or platform-policy obligations beyond technical dataset preparation.

License And Source Rights

No license identifier is declared in this package. Before public release, the publisher must add terms that they are legally authorized to grant for the audio, transcriptions, and derived uses. Public availability of source media alone does not establish permission to redistribute it or authorize commercial use. Dataset users remain responsible for complying with the final published terms and applicable source rights.

Citation

@dataset{binomar2026brahuispeech70hv2,
  author    = {Talha Bin Omar},
  title     = {BrahuiSpeech-70H V2: A Real-World Brahui Speech Recognition Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  version   = {2.0.0},
  url       = {https://huggingface.co/datasets/TBOGamer22/BrahuiSpeech-70H-V2},
  language  = {Brahui}
}

Release History

• —V2.0.0: Targeted high-disagreement audio adjudication, 1,858 corrected transcripts, 648 confirmed transcripts, and removal of one no-speech clip.
• —V1: Initial cleaned and audio-reviewed release.