Speech-to-Text
Introducing NGenSTT V2 Indic
Our most capable Speech-to-Text model, built specifically for India. Refined in TNSA’s GRACE reinforcement learning environment, it supports 27 languages, sustains accuracy in real-world noise, and transcribes code-mixed speech as it is spoken.
The model
NGenSTT V2 Indic is our flagship speech model. It is trained to be robust and general purpose, covering the full diversity of Indian accents and maintaining accuracy in noisy conditions, from crowded markets to call-centre floors.
GRACE reinforcement learning
72
hours in TNSA’s GRACE environment
Languages covered
27
22 scheduled languages, English, dialects
Model ID
ngenstt-v2-indic
available in the API Platform
Output that fits your product
Three transcription modes.
Conventional ASR systems return only a native-script transcript. NGenSTT V2 Indic returns native script, romanized text, or true code-mixed output, allowing the transcript to match the format your application expects. The same utterance, in three renderings:
Native script
Everything in the language’s own script, numerals included.
कल दोपहर तीन बजे मीटिंग है, मैंने सारी रिपोर्ट्स भेज दी हैं
Government records, publishing, subtitles
Mixed script
Native words in native script; English words and numerals in Latin.
कल दोपहर 3 बजे meeting है, मैंने सारी reports भेज दी हैं
Chat logs, support tickets, product analytics
Romanized
The full utterance in Latin script, exactly as spoken.
kal dopahar 3 baje meeting hai, maine saari reports bhej di hain
Search indexes, keyword spotting, ASCII pipelines
Capabilities
Designed for multilingual Indian speech.
01
Code-mixed speech
Speakers in India rarely use one language at a time. The model transcribes Hinglish and other mixed utterances as they are spoken, rather than mapping them onto a single language.
02
Three transcription modes
Native script for fully native output, mixed script for native words with English and numerals in Latin, and romanized for output entirely in Latin script.
03
Domain vocabulary
Extended coverage of education, agriculture, and healthcare terminology — the domains in which Indian voice products are most often deployed.
04
Language identification
Language identification is built in. The model can be used as a standalone language-ID system, or detect the spoken language and transcribe in a single pass.
Sampling rates
16 kHz studio and 8 kHz telephony
Conditions
Crowded markets, call-centre floors, open-air recordings
Accents
Indian-accented English benchmarked across 19 states
Domains
Education, agriculture, healthcare
Supported languages
27 languages, four groups.
Pass the language code at inference time, or omit it and allow the model to identify the language before transcription.
Indian-accented English
Benchmarked across speakers from 19 states.
22 constitutionally recognised languages
Full coverage of the Eighth Schedule.
Hindi dialects
Supported as distinct languages rather than Hindi variants.
Extremely low-resource
Limited public training data available.
Benchmarks
Measured on the Voice of India benchmark.
Performance is measured using Word Error Rate (WER) across 15 languages. Lower is better.
Average WER across 15 languages
WER (%), lower is better
Indic-Transcribe-core
8.7
NGenSTT V2 Indic
9.6
Saaras V3
10.7
Indic Conformer
17.8
Gemini 3 Pro
21.1
Gemini 3 Flash
23.1
Gemma E4B
36.1
OmniASR LLM 7B
44.5
OmniASR CTC 7B
62.5
Per-language results
Word Error Rate per language. The best score in each row is shown in bold; NGenSTT V2 Indic is highlighted for quick comparison.
| Language | NGenSTT V2 Indic | Indic-Transcribe-core | Saaras V3 | Indic Conformer | Gemini 3 Pro | Gemini 3 Flash | Gemma E4B | OmniASR LLM 7B | OmniASR CTC 7B |
|---|---|---|---|---|---|---|---|---|---|
| Average | 9.6% | 8.7% | 10.7% | 17.8% | 21.1% | 23.1% | 36.1% | 44.5% | 62.5% |
| Hindi | 3.4% | 3.5% | 3.8% | 6.5% | 9.3% | 7.1% | 9.1% | 9.6% | 20.7% |
| Bengali | 4.4% | 4.3% | 5.2% | 9.5% | 10.3% | 11.1% | 19.6% | 20.9% | 43.5% |
| Urdu | 5.0% | 5.0% | 7.5% | 8.0% | 10.6% | 11.0% | 14.2% | 14.8% | 89.3% |
| Marathi | 5.9% | 5.7% | 6.5% | 11.6% | 14.0% | 14.2% | 24.4% | 24.5% | 33.6% |
| Punjabi | 8.7% | 8.3% | 8.6% | 19.1% | 19.3% | 23.2% | 23.4% | 31.7% | 80.7% |
| Assamese | 8.8% | 8.2% | 9.1% | 13.1% | 23.7% | 25.0% | 45.0% | 23.9% | 33.3% |
| Odia | 9.4% | 8.7% | 11.1% | 13.1% | 25.7% | 23.3% | 44.4% | 72.3% | 90.2% |
| Gujarati | 10.0% | 9.2% | 9.7% | 16.3% | 18.1% | 20.3% | 27.3% | 32.0% | 62.5% |
| Kannada | 8.9% | 7.4% | 8.8% | 16.3% | 20.1% | 18.5% | 31.0% | 35.0% | 50.0% |
| Tamil | 9.9% | 9.0% | 9.1% | 16.2% | 15.5% | 17.3% | 37.9% | 40.6% | 58.3% |
| Telugu | 12.3% | 11.4% | 13.5% | 20.0% | 24.6% | 25.5% | 41.6% | 48.7% | 62.1% |
| Malayalam | 12.9% | 11.5% | 12.2% | 28.2% | 21.0% | 30.1% | 44.4% | 48.8% | 61.3% |
| Chhattisgarhi | 12.8% | 13.6% | 14.0% | 24.5% | 19.6% | 21.7% | 24.0% | 20.7% | 33.4% |
| Maithili | 14.2% | 11.3% | 14.2% | 16.3% | 27.2% | 28.0% | 36.7% | 44.6% | 51.6% |
| Bhojpuri | 17.6% | 13.3% | 17.9% | 30.3% | 24.1% | 20.2% | 27.0% | 26.3% | 39.2% |
WER denotes Word Error Rate; lower is better. Benchmark names and scores are presented as supplied.
Evaluation charts
Results by language group.
Grouped charts provide a closer view of the benchmark suite. Scroll horizontally to inspect each language; shorter bars indicate fewer errors.
Indo-Aryan languages
Hindi, Bengali, Urdu, Marathi, Punjabi, Assamese, Odia, and Gujarati — the highest-volume languages in Indian voice products.
Dravidian languages
Kannada, Tamil, Telugu, and Malayalam, where agglutinative morphology and long compound words raise the difficulty.
Dialects and low-resource languages
Chhattisgarhi, Maithili, and Bhojpuri, where limited public training data makes transcription hardest.
Get started