🎙️ PolyWhisper v9 — Efficient Multilingual Indic ASR

by Eulogik — Frontier Edge AI · Vernacular Intelligence · eulogik.com

TL;DR: PolyWhisper v9 is a production-ready automatic speech recognition (ASR) system for Hindi, Tamil, Telugu, Bengali, and Marathi. It pairs a frozen OpenAI Whisper-Small backbone (244M params) with tiny per-language LoRA adapters (~14MB each). Bengali WER drops −34.5% and Marathi −43.2% versus the no-augmentation baseline — at roughly 1% of the storage cost of full fine-tuning.

✨ Why PolyWhisper?

	Full fine-tune (per language)	PolyWhisper v9
Storage per language	~1.5 GB	~14 MB (100× smaller)
Backbone	retrained each time	frozen once, shared by all 5
Bengali (bn) FLEURS WER	198.8 (baseline)	130.2 (−34.5%)
Marathi (mr) FLEURS WER	170.1 (baseline)	96.7 (−43.2%)
Telugu (te) FLEURS WER	105.9 (baseline)	100.1 (−5.5%)
Hindi (hi) FLEURS WER	43.0 (baseline)	46.3
Tamil (ta) FLEURS WER	68.2 (baseline)	70.1
CPU deployment	heavy	ONNX INT8, no GPU needed

WER = word error rate (lower is better). FLEURS test set, beam=1, punctuation-normalized scoring.

📊 Benchmarks (FLEURS, beam=1, normalized WER)

Language	Code	Script	v7 (no augment)	v9 final	Δ vs v7
Hindi	hi	Devanagari	43.0	46.3	+7.7%
Tamil	ta	Tamil	68.2	70.1	+2.8%
Telugu	te	Telugu	105.9	100.1	✅ −5.5%
Bengali	bn	Bengali	198.8	130.2	✅ −34.5%
Marathi	mr	Devanagari	170.1	96.7	✅ −43.2%

🧪 The v9 finding: augment per language, not globally

Training with SpecAugment + speed perturbation on all languages damaged Hindi/Tamil (token-loop degeneration) while massively helping Bengali/Marathi. The v9 recipe augments only bn/mr and trains hi/ta clean:

Language	Augmentation	Result
Hindi, Tamil	none (clean)	matches no-augment baseline
Telugu, Bengali, Marathi	SpecAugment + 0.9×/1.1× speed perturb	large gains on hard languages

📦 Which adapter should I use?

Language	Adapter file	Backbone	WER
Hindi (hi)	polywhisper_output_hi/adapters_v3/hi_best_clean.pt	openai/whisper-small	46.3
Tamil (ta)	polywhisper_output_ta/adapters_v3/ta_best_clean.pt	openai/whisper-small	70.1
Telugu (te)	polywhisper_output_gpu0/adapters_v3/te_best_prod.pt	openai/whisper-small	100.1
Bengali (bn)	polywhisper_output_gpu0/adapters_v3/bn_best_prod.pt	openai/whisper-small	130.2
Marathi (mr)	polywhisper_output_gpu1/adapters_v3/mr_best_prod.pt	openai/whisper-small	96.7

All adapters are rank-16 LoRA (decoder + encoder attention), ~14MB each. Backbone weights are not included — they load from openai/whisper-small at runtime.

🚀 Quickstart

pip install -e .

# Hindi speech to text
polywhisper transcribe audio.wav --lang hi

# Tamil with JSON output
polywhisper transcribe audio.wav --lang ta --format json

# Auto-detect language, SRT subtitles
polywhisper transcribe audio.wav --format srt > subs.srt

# Batch a folder
polywhisper batch ./audio_folder/ --lang bn --output results.json

from polywhisper import transcribe

result = transcribe("audio.wav", lang="mr")
print(result.text)
print(result.segments)  # timestamped segments

🖥️ CPU-only inference (ONNX Runtime)

Export INT8-quantized ONNX graphs (no PyTorch, no GPU needed at inference):

polywhisper export --lang hi --variant prod --int8

Pre-exported v9 graphs live under export/onnx/ on the Hub — per language, fp32 + INT8:

Lang	Encoder (fp32 / INT8)	Decoder (fp32 / INT8)
hi	358MB / 97MB	784MB / 204MB
ta	358MB / 97MB	784MB / 204MB
te	358MB / 97MB	784MB / 204MB
bn	358MB / 97MB	784MB / 204MB
mr	358MB / 97MB	784MB / 204MB

Files are named {lang}_{lang}_best_prod_{encoder,decoder}{,_int8}.onnx. INT8 is ~4× smaller.

Verification: fp32 ONNX vs PyTorch max diff < 1e-3 on all five languages (encoder + decoder). End-to-end greedy spot-checks (FLEURS audio, beam=1):

Lang	torch WER	ONNX INT8 WER
hi (10 samples)	43.4%	48.3%
ta (5 samples)	100.0%	100.0%
te (5 samples)	100.0%	101.6%
bn (5 samples)	104.9%	118.7%
mr (5 samples)	82.9%	89.4%

Spot-checks are tiny (5–10 utterances) so single-sentence flips move the numbers; fp32 ONNX is at parity with torch. INT8 trades a few points for 4× smaller files.

🏋️ Training recipe (reproducible)

• Data: IndicVoices-ST (~19–20k clips/language) · Eval: FLEURS
• Backbone: openai/whisper-small, frozen · Adapters: LoRA rank-16, encoder + decoder attention
• Schedule: 3–5 epochs/language, batch 4, AdamW, cosine LR (peak 1e-4), 2× NVIDIA T4
• Augmentation (v9): SpecAugment + speed perturb for bn/mr only; hi/ta/te clean
• Selection: WER-gated checkpoints (*_best_*.pt) on FLEURS dev slices
• Code: train_v3.py · orchestrator kaggle_train_resumable.py · scoring normalize_ortho.py

❓ FAQ

What is PolyWhisper? PolyWhisper is an open-source Indic ASR toolkit: one frozen Whisper-Small backbone plus five small per-language LoRA adapters covering Hindi, Tamil, Telugu, Bengali, and Marathi.

How is it different from fine-tuning Whisper? Full fine-tuning rewrites ~244M–1.5B weights per language. PolyWhisper freezes the backbone and trains ~3.5M LoRA parameters per language (~14MB), so five languages ship for the storage cost of a rounding error.

Which languages are production-ready? All five ship working adapters. Hindi (46.3 WER) and Tamil (70.1) are strongest; Bengali and Marathi improved dramatically in v9 (−34.5% / −43.2% vs baseline) but remain the hardest languages.

Can I run it on CPU? Yes — export to ONNX INT8 and run with ONNX Runtime, no GPU required.

Can I run it on a Mac? Yes — PyTorch MPS is supported (Device: mps), plus CPU via ONNX.

What data was it trained/evaluated on? Trained on IndicVoices-ST conversational speech, evaluated on FLEURS read speech with punctuation-normalized, script-aware scoring.

⚠️ Limitations

• Absolute WER on Telugu/Bengali/Marathi is still high — usable for assistive/search/subtitle-draft workflows, not verbatim legal/medical transcription.
• Evaluated on read speech (FLEURS); spontaneous conversational accuracy will differ.
• Beam=1 numbers above; beam=5 decoding improves results at higher latency.

📄 License & citation

MIT. Whisper weights © OpenAI. Training data: IndicVoices-ST (CC-BY) · Eval: FLEURS (CC-BY).

@misc{polywhisper2026,
  title  = {PolyWhisper: Efficient Multilingual Indic ASR via Frozen Backbone + Per-Language LoRA},
  author = {Eulogik},
  year   = {2026},
  publisher = {HuggingFace},
  url    = {https://huggingface.co/eulogik/polywhisper}
}

🔗 Links

• 🌍 Eulogik: eulogik.com
• 🤗 Model: huggingface.co/eulogik/polywhisper
• 💻 Code: github.com/eulogik/PolyWhisper
• 🗣️ Train data: ai4bharat/indicvoices-st
• 🧪 Eval data: google/fleurs