Mr66/itantra-packs

iTantra Offline Indic Speech Packs

Speech recognition + speech synthesis for 10 Indian languages, running fully offline on a 3 GB Android phone.

Ready-to-run sherpa-onnx packs that power iTantra, a walkie-talkie that turns your voice into a 234-byte packet, hops it phone to phone over Bluetooth mesh, and speaks it aloud on the other side. No internet, no SIM, no server.

What is inside

20 packs: one speech-to-text and one text-to-speech pack per language.

Pack	Model	Size	License
<lang>-stt.itpack (9 Indic)	AI4Bharat IndicConformer, per-language, CTC head, int8	~198 MB	MIT
en-stt.itpack	NVIDIA fast-conformer CTC	~175 MB	CC-BY-4.0
<lang>-tts.itpack	Meta MMS-TTS VITS, exported with sherpa-onnx's official recipe	~114 MB	CC-BY-NC-4.0

Each .itpack is a plain zip: a manifest.json (language, kind, engine, version, sha256, sample rate) plus the ONNX model and tokens.txt. Unzip it and you can use it from any sherpa-onnx binding (Python, Kotlin, Swift, C++, JS).

Languages

Round-trip test: the TTS pack speaks a flood warning, the STT pack transcribes it back, character error rate measured (lower is better).

Language	Code	STT	TTS	Character error
Kannada	kn	Yes	Yes	1.0 %
Marathi	mr	Yes	Yes	1.2 %
Tamil	ta	Yes	Yes	2.0 %
Bengali	bn	Yes	Yes	3.2 %
Gujarati	gu	Yes	Yes	3.3 %
Telugu	te	Yes	Yes	3.3 %
Odia	or	Yes	Yes	6.0 %
Hindi	hi	Yes	Yes	8.5 %
English	en	Yes	Yes	10.2 %
Malayalam	ml	Yes	Yes	13.6 %

On a OnePlus CPH2717 (CPU only): 236 ms to decode a Hindi sentence, 644 ms to load an STT pack, 654 ms to load a voice. STT runs about 5 to 20 times faster than real time.

The error rates come from synthetic speech on a desktop CPU, not real field recordings.

Use it

In the iTantra app (easiest)

Install the APK from GitHub releases, tap Download speech models, pick your language. Done, offline from then on.

In Python

import json, zipfile
import soundfile as sf
import sherpa_onnx
from huggingface_hub import hf_hub_download

lang = "hi"
for kind in ("stt", "tts"):
    path = hf_hub_download("Mr66/itantra-packs", f"{lang}-{kind}.itpack", repo_type="dataset")
    zipfile.ZipFile(path).extractall(f"packs/{lang}-{kind}")

print(json.load(open(f"packs/{lang}-stt/manifest.json")))  # file names, engine, sample rate

Then point sherpa-onnx at the extracted files (OfflineRecognizer.from_nemo_ctc(...) for STT, OfflineTts with a VITS config for TTS). The exact file names for each pack are listed in its manifest.json.

Download everything

huggingface-cli download Mr66/itantra-packs --repo-type dataset --local-dir itantra-packs

Why these models

• Per-language STT instead of one big model. The multilingual IndicConformer needs about 2.5 GB of RAM; per-language int8 models load one at a time and fit on 3 GB phones.
• CTC head only. Fast, simple decoding with no language model, which is ideal on phone CPUs.
• One VITS voice per language. No single small multilingual TTS model covered all 10 languages at the time of building.

Licenses

The packs keep the licenses of the models they wrap:

• IndicConformer STT: MIT (AI4Bharat)
• English STT: CC-BY-4.0 (NVIDIA)
• MMS-TTS voices: CC-BY-NC-4.0 (non-commercial only) (Meta)

Please respect the non-commercial terms of the TTS voices.

Credits

AI4Bharat IndicConformer · parismitaglobalsolutions/indicconformer-sherpa-onnx · Meta MMS · k2-fsa/sherpa-onnx

Built by Team Hexabits for Smart India Hackathon 2026, PS 26173 (ISRO). If this helps you, give the dataset a like and the GitHub repo a star.