TEST REPOSITORY
TEST REPOSITORY. Models for the Flow app test builds only. Not for production or redistribution. Each model keeps its original licence (table below).
Speech models for sherpa-onnx. Each folder has its own README with source model, link and licence. catalog.json lists per folder: language, mode (streaming/offline), sherpa-onnx model type, files with size and SHA-256, url_base, licence and error per 100 words (characters for zh/ja/ko) from our 50-clip FLEURS test.
Hindi, Bengali, Marathi, Tamil, Telugu, French, Dutch and Russian are int4 weight-only compressed by us (only change). All other folders are byte-identical to the upstream sherpa-onnx downloads.
| Folder | Language | Mode | Size | Licence | Source |
|---|---|---|---|---|---|
| bn-int4b32allp | bn | offline | 84 MB | MIT | https://huggingface.co/ai4bharat/indicconformer_stt_bn_hybrid_ctc_rnnt_large |
| de-kroko-streaming | de | streaming | 68 MB | CC-BY-SA (community model) | https://huggingface.co/Banafo/Kroko-ASR |
| en-nemo-fastconformer-ctc | en | offline | 126 MB | CC-BY-4.0 | https://huggingface.co/nvidia/stt_en_fastconformer_ctc_large |
| es-moonshine-base-es | es | offline | 62 MB | NON-COMMERCIAL (Moonshine Community License) | https://github.com/moonshine-ai/moonshine |
| fr-int4b128allp | fr | offline | 67 MB | CC-BY-4.0 | https://huggingface.co/nvidia/stt_fr_fastconformer_hybrid_large_pc |
| hi-int4b128allp | hi | offline | 74 MB | MIT | https://huggingface.co/ai4bharat/indicconformer_stt_hi_hybrid_ctc_rnnt_large |
| it-nemo-10lang-transducer | it | offline | 132 MB | NVIDIA NGC Terms of Use | https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_multilingual_fastconformer_hybrid_large_pc |
| ja-moonshine-tiny-ja | ja | offline | 69 MB | NON-COMMERCIAL (Moonshine Community License) | https://huggingface.co/UsefulSensors/moonshine-tiny-ja |
| ko-moonshine-tiny-ko | ko | offline | 69 MB | NON-COMMERCIAL (Moonshine Community License) | https://huggingface.co/UsefulSensors/moonshine-tiny-ko |
| mr-int4b32allp | mr | offline | 84 MB | MIT | https://huggingface.co/ai4bharat/indicconformer_stt_mr_hybrid_ctc_rnnt_large |
| nl-int4b128allp | nl | offline | 67 MB | CC-BY-4.0 | https://huggingface.co/nvidia/stt_nl_fastconformer_hybrid_large_pc |
| ru-int4b128allp | ru | offline | 67 MB | CC-BY-4.0 | https://huggingface.co/nvidia/stt_ru_fastconformer_hybrid_large_pc |
| ta-int4b128allp | ta | offline | 74 MB | MIT | https://huggingface.co/ai4bharat/indicconformer_stt_ta_hybrid_ctc_rnnt_large |
| te-int4b128allp | te | offline | 74 MB | MIT | https://huggingface.co/ai4bharat/indicconformer_stt_te_hybrid_ctc_rnnt_large |
| zh-zipformer-ctc-small | zh | offline | 60 MB | UNVERIFIED | https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models |
| big-parakeet-v3-int4 | multi (big all-rounder) | offline | 345 MB | CC-BY-4.0 (4-bit weight-only, our change) | https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 |
| big-qwen3-asr-0.6b-int4 | multi (big all-rounder) | offline | 430 MB | Apache-2.0 (4-bit weight-only, our change) | https://huggingface.co/Qwen/Qwen3-ASR-0.6B |
| big-nemotron-3.5-streaming-int4 | multi (big all-rounder) | streaming | 359 MB | OpenMDW-1.1 (4-bit weight-only, our change) | https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b |
| lid-silero-lang95 | (95 languages) | lid | 16 MB | MIT | https://github.com/snakers4/silero-vad |
| vad-silero | (all) | vad | 1 MB | MIT | https://github.com/snakers4/silero-vad |
Notes: Moonshine (es, ja, ko) is non-commercial. Kroko is CC-BY-SA (the Kroko LICENSE file is empty; the model card states it). The NeMo 10-language model falls under the NVIDIA NGC Terms of Use. The zh model's licence could not be verified. Arabic and Whistle are not in this repo. The three big all-rounder folders (big-*) are tier "big" in catalog.json; the compressed versions of the big models are not included. it-nemo-10lang-transducer is the "europe" pack (en/de/it/es/fr/ru). Qwen3-ASR is Apache-2.0, Parakeet CC-BY-4.0, Nemotron OpenMDW-1.1.
The catalog offers only big-qwen3-asr-0.6b-int4 (447 MB); the uncompressed big-qwen3-asr-0.6b (941 MB) files stay in the repo but are not listed in catalog.json. The catalog offers only the int4 big models (big-parakeet-v3-int4, big-qwen3-asr-0.6b-int4, big-nemotron-3.5-streaming-int4). The uncompressed big-parakeet-v3, big-qwen3-asr-0.6b and big-nemotron-3.5-streaming files stay in the repo but are not listed in catalog.json.