indic-transcribe-flex

<h1 id="indic-transcribe-flex" style="color:#FFD21E;">Indic-Transcribe-flex</h1> <div align="center">

</div>

Multilingual speech recognition for 27 Indian languages — with native-script, mixed-script, and romanized output.

<div align="center"> <img src="model-diagram.png" alt="Indic-Transcribe: 27 languages in, all output features out" width="860"> </div>

Quick links — Blog · Demo · Try it out · GitHub

Indic-Transcribe-pro is a multilingual Automatic Speech Recognition (ASR) model built for 27 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.

Unlike traditional ASR systems that return only a native-script transcript, Indic-Transcribe offers three transcription modes — native script, romanized text, or true code-mixed output — so you can match the output to what your product expects.

This model is ready for commercial use.

---

<h2 id="why-choose-indic-transcribe" style="color:#FFD21E;">Why Choose Indic-Transcribe?</h2>

• 🔀 Code-mixing, natively. Indians rarely speak one language at a time. Indic-Transcribe transcribes Hinglish and other mixed speech as it is actually spoken, instead of forcing it into a single language.
• 📝 Three transcription modes. Native script for fully native output; mixed script for native words in native script with English and numerals in Latin; romanized for everything in Latin script.
• 🏥 Domain coverage where it matters. Deep vocabulary in education, agriculture, and healthcare.
• 🌐 Language identification built in. Use the model directly as a language-ID system, or let it auto-detect the language and then transcribe.
• ⚡ Streaming for live use. Real-time transcription for conversational use cases, available in Indic-Transcribe-lite.

---

<h2 id="supported-languages" style="color:#FFD21E;">Supported Languages</h2>

The model covers 27 languages across four groups:

See Supported Language Codes for the code to pass at inference time.

<h3 style="color:#FFD21E;">The three transcription modes</h3>

The same utterance, three renderings:

---

<h2 id="model-architecture" style="color:#FFD21E;">Model Architecture</h2>

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

Indic-Transcribe-pro 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 transcription mode and target language.

<h3 style="color:#FFD21E;">Feature comparison</h3>

---

<h2 id="results-at-a-glance" style="color:#FFD21E;">Results at a Glance</h2>

Results below are from the Voice of India benchmark. ASR performance is measured using Word Error Rate (WER); lower is better. Best score per column is in bold.

<h3 style="color:#FFD21E;">Systems covering all benchmark languages</h3> <!-- --- <h2 id="throughput--efficiency" style="color:#FFD21E;">Throughput & Efficiency</h2> Measured on a single NVIDIA H100, fp16. | Metric | Indic-Transcribe-pro | Indic-Transcribe-lite | | --- | :---: | :---: | | Latency | 100 ms | 30 ms | | RTFx | 500 | 1000 | | Peak GPU memory | *TBD* | *TBD* | | Checkpoint size | 4.6 GB | *TBD* | > RTFx is the inverse real-time factor: RTFx 500 means one hour of audio is transcribed in ~7.2 seconds. -->

---

<h2 id="how-to-use-this-model" style="color:#FFD21E;">How to Use this Model</h2> <h3 style="color:#FFD21E;">Installation</h3>

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.

<details> <summary>Conda environment (recommended for reproducibility)</summary>

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

</details> <h3 style="color:#FFD21E;">Input audio requirements</h3>

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

<h3 style="color:#FFD21E;">Basic inference</h3>

from huggingface_hub import snapshot_download
from indic_transcribe import IndicTranscribe

asr = IndicTranscribe.from_pretrained(snapshot_download("bodhan-ai/indic-transcribe-pro"))
print(asr("audio.wav", lang="hi"))

<h3 style="color:#FFD21E;">The three transcription modes</h3>

The same audio, three outputs. Pick the mode that matches what your downstream system expects.

Native script — everything in the language's own script. Recommended for production.

print(asr("audio.wav", lang="hi", mode="native"))
# मैंने कल पांच बजे तीन फाइलें अपलोड कीं

Mixed script (ITN) — native words in native script; English words and numerals in Latin. Inverse text normalization turns spoken numbers into digits. Recommended if you want formatted and normalized output.

print(asr("audio.wav", lang="hi", mode="mixed"))
# मैंने कल 5 बजे 3 files upload कीं

Romanized — everything transliterated into Latin script. Useful for search indexing, keyword spotting, and Latin-only UIs.

print(asr("audio.wav", lang="hi", mode="romanized"))
# maine kal 5 baje 3 files upload kin

<h3 style="color:#FFD21E;">Automatic language ID + transcription</h3>

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), ...]}

<h3 style="color:#FFD21E;">Language identification only</h3>

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: ml/ta 0.98 and kn/bn 0.96, against bho 0.05, hi 0.26, mai 0.36 and ur 0.49, each absorbed by a close neighbour. If you have a language label, pass it.

<h3 style="color:#FFD21E;">Streaming</h3>

Streaming is available in Indic-Transcribe-lite, not in pro. Use lite for live and conversational use cases where 30 ms latency matters more than peak accuracy.

---

<h2 id="inputs" style="color:#FFD21E;">Input(s)</h2>

---

<h2 id="output" style="color:#FFD21E;">Output</h2>

---

<h2 id="software-integration" style="color:#FFD21E;">Software Integration</h2>

Runtime Engine: NVIDIA NeMo (ASR collection)

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
• NVIDIA Hopper
• NVIDIA Lovelace

Supported Operating System(s): Linux

---

<h2 id="supported-language-codes" style="color:#FFD21E;">Supported Language Codes</h2>

Pass these to source_lang. Use "auto" for automatic language identification.

All 27 languages use standard ISO 639-1 / 639-3 codes (also listed in the language: field at the top of this card).

<details> <summary>Full list of language codes (27)</summary> </details>

---

<h2 id="limitations" style="color:#FFD21E;">Limitations</h2>

• 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.
• Mixed-script (ITN) mode can be wrong. Inverse text normalization is genuinely ambiguous in places, so mixed-script mode will occasionally get it wrong. For production, native-script mode is recommended, applying your own ITN downstream if you need it.
• Romanization is not standardized. There is no single canonical romanization for most Indian languages. Romanized output follows the model's learned convention, which may not match your existing transliteration scheme.
• Streaming is not available in pro. Use Indic-Transcribe-lite for real-time use.
• Low-resource languages are weaker. Bhili in particular has very little training data. 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.

---

<h2 id="license--terms-of-use" style="color:#FFD21E;">License / Terms of Use</h2>

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.

---

<h2 id="use-case" style="color:#FFD21E;">Use Case</h2>

Transcription of multilingual and code-mixed Indian-language audio, including native-script, mixed-script, and romanized output, plus language identification.

---

<h2 id="deployment-geography" style="color:#FFD21E;">Deployment Geography</h2>

Global

---

<h2 id="citation" style="color:#FFD21E;">Citation</h2>

@misc{indictranscribe2026,
  title  = {Indic-Transcribe: Multilingual Speech Recognition for 27 Indian Languages},
  author = {TODO},
  year   = {2026},
  url    = {https://huggingface.co/bodhan/Indic-Transcribe-pro}
}

---

<h2 id="ethical-considerations" style="color:#FFD21E;">Ethical Considerations</h2>

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.