The constraint
Start with the real friction.
Haryanvi and its Bangru speech variety have limited high-quality text-to-speech infrastructure, while community recordings often arrive with inconsistent metadata, duplicate content, and unlinked audio.
Speech research
Haryanvi TTS
A reproducible data and training workflow exploring natural Bangru/Haryanvi speech synthesis with IndicF5 on accessible GPU infrastructure.
- Project state
- Research in progress
- Documented
- 2026
Dataset and training workflow validated in bounded experiments; production readiness is not claimed
Bangru / Haryanvi voice lab
Sample review · 03
00:00Full precision sample00:08
01
Audit corpus
02
Gate training
03
Listen & review
Research artifact · production readiness is not claimed
Product thesis
“Regional-language voice quality starts with disciplined data, not a bigger demo.”
The response
The project builds a gated workflow around dataset reconciliation, deterministic splits, pretrained-weight validation, small overfit tests, and persisted Colab artifacts before any larger training claim is made.
System map
How the pieces fit.
01
Audit
Reconcile metadata-linked recordings, normalize transcripts, isolate rejected and orphan files, and generate review ledgers.
02
Gate
Validate model access, checkpoint structure, parameter coverage, dataset loading, and a tiny full-precision training run.
03
Train and review
Run bounded Colab experiments, persist artifacts to Drive, and listen to generated samples before considering a larger run.
What it does
Concrete capability.
- Deterministic audit of transcripts, recordings, duplicates, and orphan files
- Reproducible free-GPU training path based on a pinned IndicF5 source
- Checkpoint conversion and pretrained-parameter coverage checks
- Baseline audio, smoke training, overfit gates, and persisted review reports
Product principles
The choices behind it.
- 01Treat the metadata ledger as the authority before training
- 02Quarantine unmatched audio instead of silently adding it to the corpus
- 03Prove a tiny run can learn before spending time on scale
- 04Preserve reports, samples, and checkpoints so results can be inspected
Evidence, not theatre
What can be said clearly.
- 2,768
- metadata-linked recordings audited
- 15/15
- observed local workflow checks
- T4
- accessible Colab training target
Built with
- Python
- PyTorch
- IndicF5
- Hugging Face
- Google Colab
- FFmpeg
- Dataset audit tooling
Continue the conversation
Interested in the product—or the system behind it?
Next project
KissPDF