Bonenk/indic-transcribe-flex-GGUF

Indic-Transcribe-Flex GGUF

GGUF conversion of bodhan-ai/indic-transcribe-flex for CrispASR, using the existing Canary runtime (FastConformer encoder + Transformer decoder, 1.2B parameters).

The checkpoints keep the model's three output modes:

• native - Indic script (default)
• mixed - native script with Latin loanwords and digits (ITN)
• romanized - Latin transliteration

This is a community conversion, not an official Bodhan or NVIDIA release.

Files

File	Size	Notes
indic-transcribe-flex-f16.gguf	2.35 GiB	F16 reference
indic-transcribe-flex-q8_0.gguf	1.22 GiB	Q8_0, effectively reference quality
indic-transcribe-flex-q5_k.gguf	852 MiB	Recommended compact build
indic-transcribe-flex-q4_k.gguf	717 MiB	Smallest; one short-word regression observed on a FLEURS Hindi clip

All files include the 7,152-token aggregate SentencePiece vocabulary and the Indic prompt metadata (canary.prompt_format="indic", special token ids, and the 27-language list). No separate tokenizer download is needed.

Build CrispASR from the fork branch

The feature currently lives on the indic-transcribe-flex branch of the fork. A pull request to upstream is planned.

Linux / macOS

git clone --recursive -b indic-transcribe-flex https://github.com/Bonenk/CrispASR.git
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j"$(nproc)"

Windows

git clone --recursive -b indic-transcribe-flex https://github.com/Bonenk/CrispASR.git
cd CrispASR
.\build-windows.bat

The binary is build/bin/crispasr (build/bin/crispasr.exe on Windows). See the CrispASR README for CUDA, Metal, Vulkan, and other build options.

Run

Download one quant, then pass an explicit language with -l:

hf download Bonenk/indic-transcribe-flex-GGUF \
  indic-transcribe-flex-q5_k.gguf --local-dir .

Native script (default)

./build/bin/crispasr --backend canary \
  -m indic-transcribe-flex-q5_k.gguf \
  -f hindi.wav -l hi --canary-mode native

Mixed script (native + Latin loanwords/digits)

./build/bin/crispasr --backend canary \
  -m indic-transcribe-flex-q5_k.gguf \
  -f hindi.wav -l hi --canary-mode mixed

Romanized output

./build/bin/crispasr --backend canary \
  -m indic-transcribe-flex-q5_k.gguf \
  -f hindi.wav -l hi --canary-mode romanized

--canary-mode also accepts itn for mixed and romanised for romanized, and the environment variable CRISPASR_CANARY_MODE can set the same value.

Notes

• -l / --source-lang is required. Automatic language ID is not wired into the C++ runtime yet; the model is language-conditioned, so pass the spoken language explicitly.
• This checkpoint is ASR-only. Target language and --translate are ignored.
• Long audio is decoded in 30 s windows with 1 s overlap.
• Supported codes: as bgc bhb bho bn brx doi en gu hi hne kn kok ks mai ml mni mr ne or pa sa sat sd ta te ur (27 languages: 26 Indic languages plus Indian-accented English).

Conversion

The converter is in the fork branch at models/convert-indic-transcribe-to-gguf.py. From a checkout of the upstream HF repository:

python models/convert-indic-transcribe-to-gguf.py \
  --hf-dir /path/to/indic-transcribe-flex \
  --output indic-transcribe-flex-f16.gguf

./build/bin/crispasr-quantize \
  indic-transcribe-flex-f16.gguf indic-transcribe-flex-q5_k.gguf q5_k

The converter writes the legacy canary.* tensor layout used by CrispASR plus the Indic prompt metadata. The runtime builds the fixed 10-token prompt:

<|startofcontext|><|startoftranscript|><|emo:undefined|>
<|LANG|><|LANG|><|pnc|><noitn|itn|><|noromanized|romanized|>
<|notimestamp|><|nodiarize|>

Validation

Informal checks with CrispASR on this conversion:

• FLEURS Hindi native output matched the reference transcript exactly; mixed and romanized modes produced the expected script variants.
• FLEURS Tamil and Bengali samples produced close native transcripts; Tamil mixed mode retained Latin PH, and romanized mode produced Latin output.
• The public canary-1b-v2 JFK regression remained exact, so the shared Canary path did not regress.
• Q5_K preserved the FLEURS Hindi transcript exactly; Q4_K dropped one short word on that clip, which is why Q5_K is the recommended compact build.

These are spot checks, not a WER benchmark.

License and attribution

The source model is released under the Indic Open Model License v1.0 and is gated on Hugging Face. It also inherits the terms of its base model nvidia/canary-1b-v2, which is CC-BY-4.0. Both licenses apply to these converted weights.

You must accept the upstream license before using this repository. The license texts are included as Bodhan_AI_Open_Model_License.md and indic-open-license.md.

• Source model: bodhan-ai/indic-transcribe-flex
• Base model: nvidia/canary-1b-v2
• Runtime: CrispASR (MIT)