remiai3/STT_MODELS_FOR_APP

BharatBhasha AI - Speech-to-Text (STT) Models Catalog

📌 Repository Purpose & Motivation

This repository hosts high-efficiency, INT8-quantized ONNX acoustic models and vocabulary files for all 22 Official Scheduled Indian Languages + English.

Why this repository exists:

1. Application Independence & Resilience: Original upstream repositories are frequently gated, require authentication tokens, or face intermittent downtime. Hosting un-gated distribution endpoints guarantees that mobile and desktop clients (Android, Windows, Linux, macOS) can download models on demand without authentication barriers.
2. Bandwidth Isolation & Traffic Splitting: Decoupling Speech-to-Text (STT) from Machine Translation (T2T) and Speech Synthesis (TTS) ensures parallel high-speed downloads without rate limits or repository size bottlenecks.
3. On-Device Edge Optimization: All models are quantized to INT8 ONNX format (~188 MB each), enabling near-instant real-time transcription on edge hardware with minimal RAM overhead (< 200 MB).

---

🏗️ Repository Folder Structure

Every model and token file is placed at the repository root for clean, predictable endpoint construction:

remiai3/STT_MODELS_FOR_APP/
├── .gitattributes
├── README.md                              # Comprehensive Documentation & Model Card
├── tokens.txt                             # Unified multilingual token vocabulary
├── tokens_en.txt                          # English-specific vocabulary tokens
├── indic_conformer_tokens.txt             # Multilingual Devanagari/Indic token vocabulary
├── indic_conformer_en_tokens.txt          # English Conformer token mappings
│
├── indic_conformer_hi.onnx                # Hindi (hin_Deva) - 188 MB
├── indic_conformer_te.onnx                # Telugu (tel_Telu) - 188 MB
├── indic_conformer_bn.onnx                # Bengali (ben_Beng) - 188 MB
├── indic_conformer_ta.onnx                # Tamil (tam_Taml) - 188 MB
├── indic_conformer_mr.onnx                # Marathi (mar_Deva) - 188 MB
├── indic_conformer_gu.onnx                # Gujarati (guj_Gujr) - 188 MB
├── indic_conformer_kn.onnx                # Kannada (kan_Knda) - 188 MB
├── indic_conformer_ml.onnx                # Malayalam (mal_Mlym) - 188 MB
├── indic_conformer_pa.onnx                # Punjabi (pan_Guru) - 188 MB
├── indic_conformer_or.onnx                # Odia (ory_Orya) - 188 MB
├── indic_conformer_as.onnx                # Assamese (asm_Beng) - 188 MB
├── indic_conformer_ur.onnx                # Urdu (urd_Arab) - 188 MB
├── indic_conformer_sa.onnx                # Sanskrit (san_Deva) - 188 MB
├── indic_conformer_ne.onnx                # Nepali (npi_Deva) - 188 MB
├── indic_conformer_kok.onnx               # Konkani (gom_Deva) - 188 MB
├── indic_conformer_mai.onnx               # Maithili (mai_Deva) - 188 MB
├── indic_conformer_brx.onnx               # Bodo (brx_Deva) - 188 MB
├── indic_conformer_doi.onnx               # Dogri (doi_Deva) - 188 MB
├── indic_conformer_ks.onnx                # Kashmiri (kas_Arab) - 188 MB
├── indic_conformer_mni.onnx               # Manipuri (mni_Beng) - 188 MB
├── indic_conformer_sat.onnx               # Santali (sat_Olck) - 188 MB
├── indic_conformer_sd.onnx                # Sindhi (snd_Arab) - 188 MB
└── indic_conformer_en.onnx                # English (eng_Latn) - 166 MB

---

🌐 Direct Download Endpoints for Android & Mobile Apps

Clients can download any model file using direct HTTP GET requests without needing a Hugging Face token or API key:

1. Acoustic Model Download URL

https://huggingface.co/remiai3/STT_MODELS_FOR_APP/resolve/main/indic_conformer_{LANG_CODE}.onnx

Example for Hindi: https://huggingface.co/remiai3/STT_MODELS_FOR_APP/resolve/main/indic_conformer_hi.onnx

Example for Telugu: https://huggingface.co/remiai3/STT_MODELS_FOR_APP/resolve/main/indic_conformer_te.onnx

2. Token Vocabulary Download URL

https://huggingface.co/remiai3/STT_MODELS_FOR_APP/resolve/main/tokens.txt

---

📊 Complete Language Inventory

Language Code	Language Name	Native Script	Model Filename	Model Size	RAM Usage
hi	Hindi	हिन्दी	indic_conformer_hi.onnx	188 MB	< 200 MB
te	Telugu	తెలుగు	indic_conformer_te.onnx	188 MB	< 200 MB
bn	Bengali	বাংলা	indic_conformer_bn.onnx	188 MB	< 200 MB
ta	Tamil	தமிழ்	indic_conformer_ta.onnx	188 MB	< 200 MB
mr	Marathi	मराठी	indic_conformer_mr.onnx	188 MB	< 200 MB
gu	Gujarati	ગુજરાતી	indic_conformer_gu.onnx	188 MB	< 200 MB
kn	Kannada	ಕನ್ನಡ	indic_conformer_kn.onnx	188 MB	< 200 MB
ml	Malayalam	മലയാളം	indic_conformer_ml.onnx	188 MB	< 200 MB
pa	Punjabi	ਪੰਜਾਬੀ	indic_conformer_pa.onnx	188 MB	< 200 MB
or	Odia	ଓଡ଼ିଆ	indic_conformer_or.onnx	188 MB	< 200 MB
as	Assamese	অসমীয়া	indic_conformer_as.onnx	188 MB	< 200 MB
ur	Urdu	اُردُو	indic_conformer_ur.onnx	188 MB	< 200 MB
sa	Sanskrit	संस्कृतम्	indic_conformer_sa.onnx	188 MB	< 200 MB
ne	Nepali	नेपाली	indic_conformer_ne.onnx	188 MB	< 200 MB
kok	Konkani	कोंकणी	indic_conformer_kok.onnx	188 MB	< 200 MB
mai	Maithili	मैथिली	indic_conformer_mai.onnx	188 MB	< 200 MB
brx	Bodo	बड़ो	indic_conformer_brx.onnx	188 MB	< 200 MB
doi	Dogri	डोगरी	indic_conformer_doi.onnx	188 MB	< 200 MB
ks	Kashmiri	کٲشُر	indic_conformer_ks.onnx	188 MB	< 200 MB
mni	Manipuri	মৈতৈলোন্	indic_conformer_mni.onnx	188 MB	< 200 MB
sat	Santali	ᱥᱟᱱᱛᱟᱲᱤ	indic_conformer_sat.onnx	188 MB	< 200 MB
sd	Sindhi	سنڌي	indic_conformer_sd.onnx	188 MB	< 200 MB
en	English	Latin	indic_conformer_en.onnx	166 MB	< 200 MB

---

⚖️ Open Source Licensing & Credits

• Model Architecture & Base Weights: AI4Bharat IndicConformer by AI4Bharat (IIT Madras).
• ONNX Quantization & Runtime Compatibility: Sherpa-ONNX by Next-gen Kaldi authors.
• License: MIT License & Apache 2.0.
• Acknowledgement: Full credit goes to the AI4Bharat research team at IIT Madras and the Bhashini Mission (Ministry of Electronics and Information Technology, MeitY, Government of India) for making high-quality Indian speech AI openly accessible.