Indic-Transcribe-core

Multilingual speech recognition for 25 Indian languages, in each language's native script.

Indic-Transcribe-core is a multilingual Automatic Speech Recognition (ASR) model built for 25 Indian languages. It is trained to be robust and general purpose: it handles the full diversity of Indian accents and holds up in noisy real-world conditions, from crowded markets to call-center floors, with strong coverage in the domains where Indian voice products are actually built — education, agriculture, and healthcare.

The model transcribes into the native script of the language being spoken, and can identify the language on its own when you don't know it in advance.

This model is ready for commercial use.

---

Why Choose Indic-Transcribe-core?

• 🔀 Code-mixing, natively. Indians rarely speak one language at a time. Indic-Transcribe-core transcribes Hinglish and other mixed speech as it is actually spoken, instead of forcing it into a single language.
• 🌐 Language identification built in. Use the model directly as a language-ID system, or let it auto-detect the language and then transcribe — at the cost of one decoder step, not a second encoder pass.
• 🏥 Domain coverage where it matters. Deep vocabulary in education, agriculture, and healthcare.
• ⚡ Fast. RTFx 911 on a single H100 — one hour of audio in about four seconds.

---

Supported Languages

The model covers 25 languages across four groups:

Group	Languages
Indian-accented English	English
22 constitutionally recognised languages	Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu, Urdu
Hindi dialect	Bhojpuri
Low resource	Bhili

---

Model Architecture

Architecture Type: NVIDIA Canary — FastConformer encoder with a Transformer decoder.

Indic-Transcribe-core is built on the nvidia/canary-1b-v2 architecture. The FastConformer encoder produces acoustic representations that the Transformer decoder converts into text, with task tokens selecting the target language.

Model name	Indic-Transcribe-core
Task	Speech-to-Text (Automatic Speech Recognition)
Base model	nvidia/canary-1b-v2
Total parameters	1.2B
Encoder	FastConformer — 32 layers, 811M params, 1024 hidden dim, 8 attention heads, conv kernel 9
Decoder	Transformer — 24 layers, 419M params, 1024 hidden size, 8 attention heads
Vocabulary	7,152 tokens (1,152 special / task + 6,000 multilingual)
Sub-word algorithm	BPE (byte fallback disabled)
Precision	fp32 on disk, bf16 at inference
Checkpoint size	4.6 GB

Capabilities

Feature	Indic-Transcribe-core
Languages	25
Code-mixed audio	✅
Transcription output	Native script
Automatic language ID	✅
Inverse Text Normalization	❌
Romanized output	❌
Streaming	❌

---

Results at a Glance

Accuracy is reported as OIWER — an orthographically-informed word error rate that accepts documented spelling and transliteration variants as correct, scored against multi-reference transcripts. Lower is better.

Voice of India benchmark

283,668 clips across 14 languages.

Language	Assamese	Bengali	Bhojpuri	Gujarati	Hindi	Kannada	Maithili	Malayalam	Marathi	Odia	Punjabi	Tamil	Telugu	Urdu
OIWER	8.32	4.34	13.66	9.28	3.51	7.64	11.81	11.96	5.74	8.90	8.36	9.20	11.61	5.06

Average OIWER across the 14 languages: 8.53.

---

Throughput & Efficiency

Measured on a single NVIDIA H100, bf16, batched inference over 3,000 clips (3.11 hours of audio).

Metric	Indic-Transcribe-core
RTFx	911
Throughput	244 utterances/s
Latency (batch size 1)	286 ms median, 398 ms p90
Checkpoint size	4.6 GB

RTFx is the inverse real-time factor: RTFx 911 means one hour of audio is transcribed in about 4 seconds.

---

How to Use this Model

Installation

pip install torch torchaudio transformers sentencepiece soundfile

The model code ships inside this repository, so there is nothing else to install — no NeMo, no other toolkit.

Conda environment (recommended for reproducibility)

conda create -n indic-transcribe python=3.10 -y
conda activate indic-transcribe
pip install torch torchaudio transformers sentencepiece soundfile

Input audio requirements

Sample rate	16 kHz (resampled automatically if it differs)
Channels	Mono
Formats	.wav, .flac, .mp3
Speakers	Single speaker — see Limitations

# Convert anything to the expected format
ffmpeg -i input.mp3 -ac 1 -ar 16000 -c:a pcm_s16le audio.wav

Basic inference

from huggingface_hub import snapshot_download
from indic_transcribe import IndicTranscribe

asr = IndicTranscribe.from_pretrained(snapshot_download("bodhan-ai/indic-transcribe-core"))
print(asr("audio.wav", lang="hi"))
# मैंने कल पांच बजे तीन फाइलें अपलोड कीं

Output is always in the native script of the language being transcribed.

Automatic language ID + transcription

When you don't know the language ahead of time, omit lang. The model identifies it first, then transcribes — at the cost of one decoder step, not a second encoder pass.

text, lid = asr.transcribe("unknown_language.wav", return_lid=True)
print(lid["lang"])   # e.g. "ta"
print(text)

A language you supply always wins; identification only fills a gap. return_lid=True also works when you did supply one, so a disagreement between your metadata and the model stays visible instead of silent:

text, lid = asr.transcribe("audio.wav", lang="hi", return_lid=True)
# lid == {"lang": "hi", "source": "explicit", "topk": [("hi", 0.9999), ("ur", 0.0001), ...]}

Language identification only

To use the model purely as a language-ID system, read the predicted language and discard the transcript.

for path in ["a.wav", "b.wav", "c.wav"]:
    print(path, asr.identify(path))
    # [('ta', 0.9812), ('ml', 0.0104), ('kn', 0.0031), ...]

identify returns the ranked distribution rather than a single string, because for the confusable pairs the top-1 alone hides how close the decision was. Accuracy is uneven across languages — the Hindi-belt languages in particular are often absorbed by Hindi. If you have a language label, pass it.

You can narrow the candidate set when you know your traffic:

from indic_transcribe import RECOMMENDED_LANGS

asr.identify("audio.wav", allowed_langs=RECOMMENDED_LANGS)

This is a hard filter — audio in an excluded language is silently reassigned to the nearest permitted one rather than flagged.

---

Input(s)

Field	Details
Input Type(s)	Audio, Language ID
Input Format(s)	.wav, .flac, .mp3; string language code
Input Parameters	One-dimensional (1D) audio; one-dimensional (1D) language ID
Other Properties	16 kHz mono; audio is resampled automatically if it differs. Single speaker.

---

Output

Field	Details
Output Type(s)	Text string in the input language
Output Format(s)	String
Output Parameters	One-dimensional (1D)
Other Properties	Native-script rendering; optional detected-language tag.

---

Software Integration

Runtime Engine: PyTorch + Hugging Face Transformers (model code ships in this repository)

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
• NVIDIA Hopper
• NVIDIA Lovelace

Supported Operating System(s): Linux

---

Supported Language Codes

Pass these to lang. Omit lang for automatic language identification.

Example	Value
Hindi	lang="hi"
Tamil	lang="ta"
Bengali	lang="bn"
Indian English	lang="en"
Auto-detect	omit lang

All 25 languages use standard ISO 639-1 / 639-3 codes (also listed in the language: field at the top of this card).

Full list of language codes (25)

Language	Code	Script	Group
English (Indian)	en	Latin	Indian-accented English
Assamese	as	Bengali–Assamese	Scheduled
Bengali	bn	Bengali	Scheduled
Bodo	brx	Devanagari	Scheduled
Dogri	doi	Devanagari	Scheduled
Gujarati	gu	Gujarati	Scheduled
Hindi	hi	Devanagari	Scheduled
Kannada	kn	Kannada	Scheduled
Kashmiri	ks	Perso-Arabic / Devanagari	Scheduled
Konkani	kok	Devanagari	Scheduled
Maithili	mai	Devanagari	Scheduled
Malayalam	ml	Malayalam	Scheduled
Manipuri	mni	Bengali / Meetei Mayek	Scheduled
Marathi	mr	Devanagari	Scheduled
Nepali	ne	Devanagari	Scheduled
Odia	or	Odia	Scheduled
Punjabi	pa	Gurmukhi	Scheduled
Sanskrit	sa	Devanagari	Scheduled
Santali	sat	Ol Chiki	Scheduled
Sindhi	sd	Perso-Arabic / Devanagari	Scheduled
Tamil	ta	Tamil	Scheduled
Telugu	te	Telugu	Scheduled
Urdu	ur	Perso-Arabic	Scheduled
Bhojpuri	bho	Devanagari	Hindi dialect
Bhili	bhb	Devanagari	Low resource

---

Limitations

• Native script only. This model transcribes into the language's own script. It does not produce mixed-script/ITN or romanized output — spoken numbers stay as words, and English words inside Indic speech are rendered in the native script. Apply your own inverse text normalization or transliteration downstream if you need it.
• Single-speaker audio. The model is trained for single-speaker recordings. For multi-speaker scenarios, pair it with a diarization module and transcribe each speaker turn separately.
• Automatic language ID is uneven — do not rely on it for Hindi-belt languages. Accuracy varies widely by language, and Bhojpuri, Maithili and Urdu are absorbed by Hindi far too often to be trusted. Pass lang explicitly whenever you have it.
• Low-resource languages are weaker. Kashmiri, Maithili, Sanskrit, Bodo and Bhili have the least training data and the highest error rates. Check the per-language numbers in Results at a Glance before committing to a language.
• Noise robustness has limits. Heavily overlapped speech, far-field audio, and very low-SNR recordings will still degrade quality.
• Streaming is not available. This model is for batch and request-response transcription.

---

License / Terms of Use

Released under BODHAN AI OPEN MODEL LICENSE 1.0.

The base model, nvidia/canary-1b-v2, carries its own license terms — ensure your use complies with both.

---

Use Case

Native-script transcription of multilingual and code-mixed Indian-language audio, plus language identification.

---

Deployment Geography

Global

---

Citation

@misc{indictranscribe2026,
  title  = {Indic-Transcribe: Multilingual Speech Recognition for Indian Languages},
  author = {TODO},
  year   = {2026},
  url    = {https://huggingface.co/bodhan-ai/indic-transcribe-core}
}

---

Ethical Considerations

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Developers should work with their team to ensure this model meets requirements for the relevant industry and use case, and addresses unforeseen product misuse.