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.

• 🔀 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.

The model covers 25 languages across four groups:

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.

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.

Indic-Transcribe Core offers better accuracy. Do check out Indic-Transcribe Flex for support for more transcription modes and languages.

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

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

# Gated model: accept the terms on this page, then run `hf auth login` once (older installs: `huggingface-cli login`).
import sys
from huggingface_hub import snapshot_download

model_dir = snapshot_download("bodhan-ai/indic-transcribe-core")
sys.path.insert(0, model_dir)          # the model code ships inside the download

from indic_transcribe import IndicTranscribe
asr = IndicTranscribe.from_pretrained(model_dir)
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.

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), ...]}

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.

asr(...) decodes a file in one pass and refuses audio longer than 45 s. That limit is deliberate: the model trains on clips of up to 30 s, and a single pass over long audio collapses into repetition rather than degrading gracefully. For long audio, use the chunked helper that ships with the model. It cuts the audio at natural pauses (or at exactly 30 s when there is none) into pieces of at most 30 s, transcribes each, and joins the text:

from long_form import transcribe_long

print(transcribe_long(asr, "long_audio.wav", lang="hi"))

Omit lang and the language is identified once, from samples across the whole file. Pass it whenever you know it — it is faster and avoids a wrong guess.

How the long-audio path works:

audio (.wav / .flac / .mp3, any sample rate, mono or stereo)
  │  load → mono → 16 kHz
  ▼
lang given? ── no ──► identify it (long files: 3 × 20 s windows across the file)
  │
  ▼
≤ 30 s? ── yes ──► decode in one pass ───────────────────────────┐
  │ no                                                           │
  ▼                                                              │
split at pauses                                                  │
  • pause = ≥ 250 ms at least 25 dB below the file's peak        │
  • cut in the middle of the pause nearest 25 s                  │
  • no pause within 30 s → cut at exactly 30 s                   │
  ▼                                                              │
decode each piece (every piece ≤ 30 s)                           │
  • output looping? → split that piece at a pause, decode again  │
  • still looping?  → collapse the repeated phrase               │
  ▼                                                              │
join the pieces ─────────────────────────────────────────────────┴──► transcript

Pieces are kept as long as possible because every cut costs a little accuracy — the words at a cut are the ones most likely to be wrong — and 30 s is the longest input the model was trained on.

inference.py ships with the model and wraps the same calls. It picks the whole-file or chunked path by duration, and identifies the language when --lang is omitted.

MODEL_DIR=$(python -c 'from huggingface_hub import snapshot_download; print(snapshot_download("bodhan-ai/indic-transcribe-core"))')

python "$MODEL_DIR/inference.py" audio.wav --lang hi
python "$MODEL_DIR/inference.py" long_call.wav --show-lang      # no --lang: identify it, and print it
python "$MODEL_DIR/inference.py" *.wav --lang ta                # several files

nemo/ holds the same model as an NVIDIA NeMo checkpoint, for pipelines already built on NeMo. Its config names the tokenizer class the model was trained with, which is not part of a stock NeMo install, so load_nemo registers the copy that ships beside it before restoring:

import sys
from huggingface_hub import snapshot_download

model_dir = snapshot_download("bodhan-ai/indic-transcribe-core")
sys.path.insert(0, f"{model_dir}/nemo")      # the loader ships inside the download

from load_nemo import load_nemo_model
model = load_nemo_model(model_dir)
print(model.transcribe(["audio.wav"], source_lang="hi", target_lang="hi", pnc="yes")[0].text)

Nothing inside NeMo is modified, and the loader finds the checkpoint next to itself. If you would rather not add anything to sys.path, copy nemo/canary_multilingual_tokenizer.py into your NeMo install beside canary_tokenizer.py and restore_from the .nemo directly.

Give it a language. Pass any of the codes in Full list of language codes below as both source_lang and target_lang:

as  bho  bn  brx  doi  en  gu  hi  kn  kok  ks  mai  ml  mni
mr  ne  or  pa  sa  sat  sd  ta  te  ur

The checkpoint keeps one shared sub-tokenizer for all of them, and the class above routes every code to it. Without it, stock NeMo raises RuntimeError: Unsupported language: 'ml'. This checkpoint writes each language in its native script.

pip install "nemo_toolkit[asr]"

MODEL_DIR=$(python -c 'from huggingface_hub import snapshot_download; print(snapshot_download("bodhan-ai/indic-transcribe-core"))')

python "$MODEL_DIR/nemo/test_nemo.py" audio.wav hi          # check the install
python "$MODEL_DIR/nemo/inference_nemo.py" audio.wav --lang hi
python "$MODEL_DIR/nemo/inference_nemo.py" long_call.wav --lang ta   # chunked automatically
python "$MODEL_DIR/nemo/inference_nemo.py" *.wav --lang bn --batch-size 8

Audio over 30 s is cut at natural pauses into pieces of at most 30 s and decoded piece by piece -- the same splitter, cap and repetition guard the Hugging Face model uses, so both paths cut audio identically. inference_nemo.py does it automatically; in code:

from long_form_nemo import transcribe_long_nemo

print(transcribe_long_nemo(model, "long_audio.wav", lang="hi", batch_size=8))

Batching the pieces makes this path the faster one for long audio: a 10-minute file takes 8 s at batch_size=8 against 24 s at 1, and an hour of audio runs in 48 s (77x realtime, 9.2 GiB of GPU memory). Pass lang -- this path has no language identification of its own, and a wrong language gives you the wrong script rather than an error.

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

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

• 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.

Released under Indic Open Model License v1.0.

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

If you find the license difficult to understand, here is a plain-language guide to the Indic Open Model License.

Broad, no-cost access for research, government, nonprofit, and commercial use — with a few conditions attached.

This deed is a human-readable summary of the license, not a substitute for it. Where the two disagree, the full Indic Open Model License governs.

You're free to

No cost, no royalty, worldwide — for research, government, nonprofit, and commercial use, at any scale.

• ✅ Run it — for inference, in a product, in research, however you like.
• ✅ Change it — fine-tune, distill, quantize, merge, or otherwise build on it.
• ✅ Self-host it — power your own product or service with it, commercial or not.
• ✅ Share it — pass on copies of the model or your own version of it.

As long as you

Five conditions cover almost everything. The rest of the license is these, spelled out in legal detail.

1. Give credit

Wherever you ship the model or a derivative to anyone else, say where it came from — and don't strip out existing notices.

"Built with [Model Name] from Bodhan AI / AI4Bharat."

2. Pass it on the same way

If you give your fine-tuned or derived version to anyone else — hand it over, or run it as a service for them — it carries this exact license. You can't relicense it on different terms.

3. Ask before hosting it for others

Self-hosting is free. But if you're going to run it as an API or hosted service that other people or companies call directly, that needs Bodhan AI's written sign-off first — unless you're a nonprofit, government, or academic user, or you publicly release an equally capable open version within 90 days.

4. Don't use it to cause harm

No exceptions — not even for nonprofit or research use. That means no:

• child sexual abuse material, or content that sexualizes minors
• weapons development, including chemical, biological, radiological, or nuclear
• mass surveillance or social-scoring systems
• disinformation campaigns, including election manipulation
• automated decisions that affect someone's legal rights without human oversight
• deepfakes or voice clones of real people without their consent
• robocalls, auto-dialers, or voice-phishing scams
• AI companion products designed to simulate romance or foster emotional dependency

5. Talk to us if your product gets huge

If your own product built on this — not through hosting it for others, that's covered above — crosses either threshold, you'll need a separate commercial license. Doesn't apply to nonprofit, government, or academic users.

Good to know

• No warranty. The model is provided as-is. It isn't tested or certified for safety-critical use — medical, aviation, nuclear, or similar — so test thoroughly before relying on it in high-stakes settings.
• You handle your own compliance. Export controls, sanctions, and data-protection law (including India's DPDP Act, where it applies) are on you, not Bodhan AI.
• This deed doesn't replace the license. It leaves out most of the legal detail — termination, dispute resolution, confidentiality, and more all live in the full text. Read that before you rely on anything here.

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

Global

@misc{indictranscribe2026,
  title  = {Indic-Transcribe: Built for the way India actually speaks},
  author = {Bodhan AI, AI4Bharat},
  year   = {2026},
  url    = {https://bodhan.ai/research/blogs/indic-transcribe}
}

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.