Dataset Card for Bangla Speech Corpus v1.0 (bangla_speech_corpus_v1.0)

Mandatory Human Audit & Quality Disclaimer

"This corpus is an automatically quality-filtered Bangla speech corpus with a stratified 200-sample human audit."

"Only 200 samples were manually audited. The remaining samples were not manually verified."

This dataset is not "fully human verified", "completely manually validated", or "100% human reviewed". Quality guarantees rest upon rigorous, deterministic automated acoustic and text filters validated against a stratified 200-sample human audit across diverse acoustic strata.

---

1. Dataset Summary

Bangla Speech Corpus v1.0 is an open-benchmark Bengali speech corpus engineered for acoustic modeling, automatic speech recognition (ASR), and voice adaptation. Derived from authentic televised broadcasts and educational panel discussions across 6 major networks, the dataset provides 24 kHz high-fidelity linear PCM audio paired with authentic closed captions.

Transcripts are sourced strictly from broadcast closed captions without synthetic text generation, paraphrasing, or external automated speech recognition models. All numerals are deterministically verbalized into standard written Bengali words to ensure consistency for downstream acoustic and language modeling.

---

2. Dataset Version

• Dataset Version: 1.0.0
• Release Date: 2026-09-30
• Integrity Status: TRAINING_READY_WITH_PARTIAL_HUMAN_AUDIT

---

3. Language

• Language: Bengali / Bangla (bn, bn-BD)
• Script: Bengali Unicode (ঀ–৿)

---

4. Task

• Primary Task: Automatic Speech Recognition (ASR)
• Secondary Tasks: Text-to-Speech (TTS) acoustic adaptation, Voice Activity Detection (VAD), Speech Quality Assessment.

---

5. Dataset Statistics (Public Candidate Partitions)

Partition	Segments	Total Duration	Duration (Hours / Mins)	Independent Sources	Distinct Speakers	Role
train	3,813	28,668.00 s	7.96 hrs (477.80 min)	44	44	Primary acoustic model training
validation	232	1,586.50 s	0.44 hrs (26.44 min)	4	4	Hyperparameter tuning & model selection
eval	305	2,406.60 s	0.67 hrs (40.11 min)	2	2	Held-out out-of-domain evaluation
Total Public Candidate	4,350	32,661.10 s	9.07 hrs (544.35 min)	50	50	Public benchmark corpus

(Note: The 2-segment, 21.55-second cryptographically locked regression test set test.csv is strictly private and is NOT included in this public candidate package).

---

6. Audio Format

• Sampling Rate: 24,000 Hz
• Channels: 1 (Mono)
• Bit Depth: 16-bit Linear PCM (PCM_16)
• Format: Uncompressed WAV
• Duration per Segment: $2.00\text{ s} \le t \le 15.00\text{ s}$ (Mean: $7.51\text{ s}$)
• Peak Signal Level: $\le -1.00\text{ dBFS}$ (Zero clipping)
• Signal-to-Noise Ratio (SNR): Mean $23.55\text{ dB}$ (WADA-SNR algorithm; quality threshold $\ge 10.0\text{ dB}$)

---

7. Data Collection

Audio and closed captions were captured from public broadcast news, panel discussions, interviews, and educational lectures across six prominent Bangladeshi broadcast television networks:

1. BBC News Bangla (bbc_bangla)
2. Deutsche Welle Bangla (dw_bangla)
3. Jamuna TV (jamuna_tv)
4. Somoy TV (somoy_tv)
5. Channel 24 (channel24)
6. BanglaVision (banglavision)
7. Educational Lectures (educational)

---

8. Transcript Source

Transcripts originate exclusively from broadcast closed captions (YouTube CC).

• Zero Synthetic Generation: Transcripts are never generated, hallucinated, or rephrased by Large Language Models (LLMs).
• Zero External ASR Dependency: No Whisper, Faster-Whisper, Google Cloud Speech-to-Text, Sarvam STT, or proprietary ASR service was used to produce, alter, or align transcripts.

---

9. Processing Pipeline

1. Acoustic Ingestion: Demuxing and conversion to canonical 24,000 Hz single-channel 16-bit PCM WAV.
2. Segmentation & VAD: Splitting audio into bounded clips ($2.0\text{s} - 15.0\text{s}$) with speech activity ratio $\ge 0.75$.
3. Acoustic Quality Scoring: Automated multi-criteria scoring assessing SNR ($\ge 10\text{ dB}$), clipping, speaking rate ($5 - 25\text{ chars/sec}$), and overall quality score ($Q \ge 75.0$).
4. Text Standardization: Punctuation standardization, non-speech tag stripping ([মিউজিক], [হাসি]), Unicode NFC normalization.
5. Numeral Verbalization: Deterministic rule-based verbalization into spoken Bengali words.

---

10. Speaker Handling

Single-speaker attribution is enforced for clean training partitions:

• Each segment in train and validation is associated with a verified individual source recording and distinct speaker identifier (speaker_id).
• Across the public candidate partitions, 50 independent broadcast programs and distinct speakers are represented.

---

11. Multi-Speaker Filtering

In broadcast panel talk shows, closed captions frequently record dialogue exchanges with leading >> indicators denoting turn handoffs.

• To prevent assigning multi-speaker speech to single speaker profiles, all 1,163 segments containing conversational turns (>>) in clean transcripts were permanently isolated and excluded from single-speaker training splits.
• Observed Contamination in Clean Partitions: 0 multi-speaker segments.

---

12. Numeral Normalization

Digit tokens (০-৯ and 0-9) present severe out-of-vocabulary challenges for acoustic modeling. A deterministic, rule-based normalizer (scripts/numeral_normalizer.py) converts all numbers into standard written Bengali words:

• Cardinals (0–99): Complete dictionary mapping (e.g. ৮৪ $\to$ "চুরাশি").
• Place-Value System: Hundreds (শত), thousands (হাজার), lakhs (লাখ), and crores (কোটি).
• Decimals & Percentages: ৩.৮৮ $\to$ "তিন দশমিক আট আট", ১০০% $\to$ "একশত শতাংশ".
• Date Ordinals: ১লা $\to$ "পয়লা", ২৫শে $\to$ "পঁচিশে".
• Clitic Preservation: Classifiers (১০টি $\to$ "দশটি", ২০জন $\to$ "বিশজন") are preserved with morphological integrity.

---

13. Alignment & CTC Boundary Pruning

Segment boundaries were initially established using broadcast timestamp markers verified against Voice Activity Detection (VAD) audio envelopes. Any segment where speech extended beyond the boundary window or severed dependent clauses was isolated into a review manifest and excluded from training.

Acoustic CTC Forced Alignment Boundary Pruning

Because broadcast closed captions often utilize rolling temporal buffers (~4.5-second snippets with ~2-second overlaps), raw caption segment cuts frequently introduce non-spoken boundary words at snippet heads and tails.

To ensure pure phonetic correspondence between speech and text:

• Acoustic Model: arijitx/wav2vec2-xls-r-300m-bengali (50 Hz frame emissions, 20 ms hop).
• Alignment Method: torchaudio.functional.forced_align and token-merging.
• Punctuation Alignment: Non-Bengali punctuation stripped during alignment tokenization to prevent out-of-vocabulary penalty, while retaining surface punctuation in transcripts.
• Boundary Pruning Threshold: Calibrated acoustic log-probability threshold of -6.5. Edge words forced into silence score well below -6.5 and are iteratively pruned until reaching authentic speech.
• Coverage: Applied comprehensively across all corpus splits (train, validation, and eval). Over 92% of segments were purified of caption boundary bleed.

---

14. Quality Filtering

All included segments meet the following strict thresholds:

• Minimum Duration: 2.00 s
• Maximum Duration: 15.00 s
• Minimum WADA-SNR: 10.0 dB
• Maximum Peak Amplitude: -1.0 dBFS
• Minimum Speech Activity Ratio: 0.75
• Minimum Quality Score: 75.0 / 100.0

---

15. Human Audit

A stratified forensic human audit was conducted on 200 representative audio segments across four strata:

• Stratum 1 (Numerals, 51 segments): Verified 100% verbalization accuracy (68 of 68 numeral instances correctly verbalized).
• Stratum 2 (Formatting-only >>, 31 segments): Verified that leading broadcast turn symbols were cleanly removed with audio containing only a single speaker.
• Stratum 3 (Rejected Multi-Speaker, 52 segments): Confirmed 100% precision of automated multi-speaker rejection; every sample contained interleaved dialogue.
• Stratum 4 (General Distribution, 66 segments): Confirmed high acoustic clarity and transcript fidelity across diverse television sources.

---

16. Dataset Splits

• train (train_frozen.csv): 3,813 segments (7.96 hrs, 44 sources, 44 speakers)
• validation (validation.csv): 232 segments (26.44 min, 4 sources, 4 speakers)
• eval (eval_test.csv): 305 segments (40.11 min, 2 sources, 2 speakers)
• Total Public Candidate: 4,350 segments (9.07 hrs)

---

17. Leakage Prevention

To guarantee valid acoustic and language model evaluation, strict separation is enforced:

• Train / Evaluation Source Overlap: 0 sources (Completely source-disjoint).
• Train / Evaluation Speaker Overlap: 0 speakers (Completely speaker-disjoint).
• Train / Evaluation Transcript Overlap: 0 sentences (Zero sentence duplication between train and evaluation).
• Multi-Speaker Contamination in Clean Partitions: 0 segments.

---

18. Limitations

1. Partial Human Verification: Only 200 samples were manually audited; the remaining 4,150 segments were verified solely by deterministic programmatic filters.
2. Broadcast Acoustic Bias: The recordings reflect broadcast television studio and field news environments, which may contain controlled studio reverberation, lapel microphone characteristics, or subtle background ambience.
3. Dialect Coverage: Speech primarily reflects Standard Colloquial Bengali (SCB) as used in national news media and formal academic discourse. Regional dialects (e.g. Chittagonian, Sylheti, Noakhali) are underrepresented.

---

19. Intended Use

• Academic research in Bengali speech recognition (ASR) benchmarking.
• Acoustic phoneme model training and feature extraction.
• Single-speaker voice adaptation and Text-to-Speech (TTS) acoustic modeling.

---

20. Out-of-Scope Use

• Commercial deployment or redistribution without explicit licensing from original television broadcast copyright holders.
• Speaker biometric identification or surveillance applications.
• Synthetic voice impersonation without consent.

---

21. License & Rights Notice

Redistribution and licensing of the underlying third-party broadcast audio must be verified before public release.

• The audio recordings originate from third-party public television and educational broadcasts across Bengali networks.
• Audio content was compiled and quality-filtered strictly under the doctrine of Academic Research Fair Use for computational linguistics and automatic speech recognition benchmark research.
• No public redistribution license (such as Creative Commons, MIT, or Apache 2.0) has been granted by the original broadcast copyright holders.
• This dataset is NOT licensed under CC-BY, CC-BY-SA, MIT, Apache-2.0, or CC0.

---

22. Citation

@dataset{bangla_speech_corpus_v1,
  author       = {Hasan, Ragib and Sen Toma, Indrani and Galib, Asadulla Hel},
  title        = {Bangla Speech Corpus v1.0},
  version      = {1.0.0},
  year         = {2026},
  publisher    = {Hugging Face},
  doi          = {10.57967/hf/10781},
  url          = {https://huggingface.co/datasets/ragibhasan/bangla_speech_corpus_v1}
}