Source-linked AI summary
BanglaKontho: Closing the Long-Form Gap in Bangla Text-to-Speech
Mizbaul Haque Maruf
TL;DR
Bangla TTS lacks high-quality data for long-form prosody, as public resources are dominated by short read-prompt recordings. BanglaKontho pairs a 20-hour single-speaker audiobook corpus with verified transcripts and reusable normalization tooling; its baseline reports lower WER and higher naturalness MOS than the same architecture retrained on IndicTTS-Bn.
Problem
Method
BanglaKontho provides a 20-hour single-speaker audiobook corpus with verified transcripts, a reusable text normalizer, and a reproducible preprocessing pipeline.
Source: §1 · p. 1, §3.5 · p. 2Results
9.5% WER and 4.46 naturalness MOS, versus 16.0% and 3.16 for the same architecture retrained on IndicTTS-Bn; MCD is comparable (5.85 vs. 5.97).
Source: §4.2 · p. 3Takeaways & Limitations
The resource addresses the long-form Bangla speech-data gap and provides ready-to-use tooling for the broader Indic-TTS community.
Source: §6 · p. 4The corpus is single-speaker and audiobook-style; conversational use is recommended with a small in-domain adaptation set.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Bangla, the seventh most spoken language in the world, remains under-resourced for neural text-to-speech. Public Bangla speech corpora are dominated by short read-prompt utterances collected for speech recognition, leaving long-form prosody and consistent single-speaker narration uncovered. We present BanglaKontho, a single-speaker Bangla TTS corpus of 20 hours derived from professional audiobook recordings: 7,050 segmented utterances with verified transcripts at 24 kHz. We also release a reusable Bangla text normalizer covering Bangladeshi-style digit grouping, currency and date expressions, Danda punctuation and Unicode normalization, together with the full preprocessing pipeline. An MB-iSTFT-VITS baseline trained from scratch reaches 9.5% WER and 4.46 naturalness MOS, against 16.0% and 3.16 for the same architecture retrained on the 12-hour IndicTTS-Bn corpus. The corpus is released openly under CC BY-NC 4.0.
1. Introduction
Bangla TTS lacks high-quality data suited to long-form prosody, while existing resources are dominated by short read-prompt recordings. BanglaKontho contributes an audiobook corpus, reusable normalization tooling, and a reproducible baseline.
1. Introduction
- Bangla TTS resources are scarce and fragmented, with short read-prompt recordings leaving long-form rhythm and sentence-level intonation under-covered.These qualities matter for applications including audiobook synthesis, screen readers, and conversational agents.Source: §1 · p. 1
- BanglaKontho is a 20-hour single-speaker audiobook corpus with verified transcripts, designed to complement short read-prompt Bangla TTS corpora.It is released under CC BY-NC 4.0.Source: §1 · p. 1
- The release includes a Bangla text normalizer and reproducible end-to-end preprocessing pipeline, alongside an MB-iSTFT-VITS baseline.The baseline is benchmarked against IndicTTS-Bn and a normalizer-disabled ablation.Source: §1 · p. 1
2. Related work
Prior Bangla speech resources address different needs but do not combine a trained single speaker, long-form audiobook prosody, studio-clean audio, and verified transcripts. The related work also identifies gaps in Bangla-specific normalization tools and released audiobook data.
Bangla speech resources
- Existing Bangla speech resources vary in scale and purpose, but none combines a single trained speaker, long-form audiobook prosody, studio-clean audio, and verified transcripts.The cited landscape includes crowd-sourced read-prompt corpora, IndicTTS, and weakly aligned Shrutilipi.Source: §2 · p. 1
Bangla TTS
- Prior Bangla TTS spans statistical-parametric systems and neural voice cloning, while this work adopts a VITS-based architecture [3] for comparison and keeps the corpus architecture-agnostic.The passage notes that the voice-cloning approach relies on centralized fine-tuning of raw speech.Source: §2 · p. 1
Text normalization for Indic scripts
- Bangla normalization must handle non-compositional number words, Unicode-distinct graphemes, and mixed Bangla and ASCII digits; prior work noted the need but released no reusable tool.BanglaKontho addresses this tooling gap.Source: §2 · pp. 1–2
3. Dataset construction
The dataset pipeline converts professional audiobook recordings into sentence-level clips with manually reviewed transcripts and reusable Bangla normalization. It uses separate analysis and release audio representations, then applies speaker verification and quality filtering.
3.1. Audio acquisition and preprocessing
- The analysis copy is 16 kHz mono PCM for segmentation and speaker verification, while release clips are cut from native-rate audio and resampled to 24 kHz.Post-production dynamic-range compression is retained rather than inverted.Source: §3.1 · p. 2
3.2. Speaker verification
- Speaker verification uses pyannote.audio voice activity detection and ECAPA-TDNN embeddings, discarding segments below a calibrated cosine-similarity threshold against the enrolled narrator.Source: §3.2 · p. 2
3.3. Segmentation
- VAD-derived silence boundaries guide segmentation toward likely sentence ends; clips under 1.0 s or over 20 s are dropped, with a 2–15 s operating target.Source: §3.3 · p. 2
3.4. Transcription and manual verification
- Two native Bangla volunteers independently reviewed every clip after first-pass Gemini transcription; agreement was κ = 0.91 on a 500-clip overlap subset.Gemini achieved 2.7% WER against author-transcribed references on a random 500-clip subset, serving as a starting point rather than ground truth.Source: §3.4 · p. 2
3.5. Bangla text normalization
- Bangla normalization composes Unicode forms, verbalizes symbols and numbers, handles lakh/coti grouping, decimals and dates, and offers a strict Bangla-block mode.The module also normalizes Danda punctuation.Source: §3.5 · p. 2
3.6. Quality filtering and statistics
- After transcription and normalization, clips are dropped for residual non-speech tails, anomalous duration-to-text ratios, or verifier flags.Final statistics are reported in Table 2.Source: §3.6 · p. 2
4. Baseline TTS and evaluation
This section evaluates an MB-iSTFT-VITS baseline on BanglaKontho against a retrained IndicTTS-Bn model and natural speech, then tests the contribution of Bangla text normalization. The BanglaKontho model has lower WER and higher naturalness MOS than the IndicTTS-Bn baseline, while disabling normalization worsens WER and MOS.
4. Baseline TTS and evaluation
- The from-scratch MB-iSTFT-VITS baseline uses 22 050 Hz training audio from the 16 kHz analysis copy, rather than the 24 kHz release.The configuration has 34.7 M parameters and uses a multi-band iSTFT decoder with PQMF subband synthesis.Source: §4 · p. 2, Figure 1 · p. 3
4.1. Evaluation protocol
- Evaluation uses 472 chapter-disjoint test utterances, IndicWav2Vec WER, DTW-aligned MCD, and naturalness MOS ratings from 15 native Bangla raters.Raters scored 50 length-balanced utterances per system; natural test-split audio yields an 8% WER ASR floor.Source: §4.1 · p. 3, Table 2 · p. 3, Table 3 · p. 3
4.2. Baseline results
- 9.5% vs. 16.0% WER and 4.46 vs. 3.16 naturalness MOS; MCD is comparable at 5.85 vs. 5.97 for BanglaKontho and IndicTTS-Bn, respectively.The BanglaKontho model approaches the natural-speech ceiling of 4.90 MOS and 8% ASR-floor WER.Source: Table 3 · p. 3, §4.2 · p. 3
4.3. Normalizer ablation
- Disabling bangla_normalize yields WER 14.0%, naturalness MOS 3.80, and comparable MCD 5.81; the gap concentrates on symbol-bearing utterances.The affected utterances contain numerals, dates, currency symbols, or mixed Bangla/ASCII digits; these make up 18% of the test set, and the gap narrows on symbol-free utterances.Source: §4.3 · pp. 3–4
5. Limitations and ethical considerations
The corpus is a single-speaker audiobook resource, with narrative prosody that differs from conversational TTS. Its release is restricted to non-commercial academic research and does not claim cryptographic speaker anonymity.
Scope
- The single-speaker audiobook scope produces narrative-reading prosody; conversational TTS use is recommended to include a small in-domain adaptation set.Multi-speaker extension is left to future work.Source: §5 · p. 4
License and release
- The corpus is released under CC BY-NC 4.0 for non-commercial academic research, with commercial use and specified harmful uses prohibited by the data-use agreement.Rights holders may request takedowns, and corresponding clips will be removed.Source: §5 · p. 4
Speaker anonymity and ASR-based intelligibility
- Omitting the narrator’s identity from metadata does not prevent voice-based identification by people with the audio; the paper makes no cryptographic anonymity claim.WER also depends on IndicWav2Vec accuracy, so synthesis scores are reported relative to the 8% natural-speech ASR floor.Source: §5 · p. 4
6. Conclusion
The conclusion presents BanglaKontho as a single-speaker audiobook corpus paired with a reusable normalizer and MB-iSTFT-VITS baseline. It identifies multiple narrators, preference testing, and downstream application evaluation as future work.
Contribution
- BanglaKontho provides a 20-hour single-speaker Bangla audiobook corpus, a reusable text normalizer, and an MB-iSTFT-VITS baseline.The paper frames the resource as addressing a gap in long-form Bangla speech data and providing tooling for the Indic-TTS community.Source: §6 · p. 4
7. Generative AI Use Disclosure
Generative AI was used for first-pass clip transcripts and manuscript language editing under author responsibility. The authors state that no scientific content, experimental design, analysis, or conclusions were produced by an AI tool.
Transcript generation
- Gemini 2.0 Flash-Lite generated first-pass transcripts, but no transcript entered the release without independent verification by two native Bangla speakers.First-pass transcription quality was 2.7% WER on a hand-transcribed subset.Source: §7 · p. 4
Manuscript editing
- Generative AI assistants were also used for manuscript language editing and formatting.Source: §7 · p. 4
Author responsibility
- The authors state that AI tools produced none of the scientific content, experimental design, analysis, or conclusions, and that the author takes responsibility.Source: §7 · p. 4