Fun-ASR-Nano-2512-GGUF/README.md
| 1 | --- |
| 2 | library_name: audio.cpp |
| 3 | license: other |
| 4 | license_name: funasr-model-license-1.1 |
| 5 | license_link: https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512-hf/blob/main/README.md |
| 6 | base_model: FunAudioLLM/Fun-ASR-Nano-2512-hf |
| 7 | pipeline_tag: automatic-speech-recognition |
| 8 | tags: |
| 9 | - audio.cpp |
| 10 | - gguf |
| 11 | - speech-recognition |
| 12 | - multilingual |
| 13 | - funasr |
| 14 | --- |
| 15 | |
| 16 | # Fun-ASR-Nano-2512 GGUF |
| 17 | |
| 18 | Standalone audio.cpp GGUF builds of |
| 19 | [FunAudioLLM/Fun-ASR-Nano-2512-hf](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512-hf). |
| 20 | Each file embeds the model configuration, processor configuration, tokenizer, |
| 21 | chat template, and the audio.cpp model package specification. |
| 22 | |
| 23 | ## Files |
| 24 | |
| 25 | | File | Size | SHA256 | |
| 26 | | --- | ---: | --- | |
| 27 | | `fun-asr-nano-2512-q8_0.gguf` | 1,045,334,432 bytes | `4d727357574b079b7f43336b2930f39da086ca02f5d8d50872090b4c1c3d5e0a` | |
| 28 | | `fun-asr-nano-2512-f16.gguf` | 1,675,708,832 bytes | `3d906c3ccfed07efef88ff53d6cc94b788b9d2edf1492a5679d041b43e98c5be` | |
| 29 | |
| 30 | The source checkpoint is pinned to revision |
| 31 | `854d88f94205cd17d2afdb24332130d86fbe654a`. The source |
| 32 | `model.safetensors` SHA256 is |
| 33 | `335ca3e74917f1156690400e2c344350112950165789cf78ce3d0a367affd821`. |
| 34 | |
| 35 | ## audio.cpp |
| 36 | |
| 37 | ```bash |
| 38 | audiocpp_cli \ |
| 39 | --task asr \ |
| 40 | --family fun_asr_nano \ |
| 41 | --model fun-asr-nano-2512-q8_0.gguf \ |
| 42 | --backend cuda \ |
| 43 | --audio speech.wav |
| 44 | ``` |
| 45 | |
| 46 | Fun-ASR-Nano currently provides offline multilingual ASR. It does not expose |
| 47 | streaming or timestamp output. On CUDA, audio.cpp keeps the Q8_0 encoder and |
| 48 | adaptor weights native and loads decoder weights as BF16 by default for stable |
| 49 | logits. An explicit `fun_asr_nano.decoder_weight_type` session option overrides |
| 50 | that default. |
| 51 | |
| 52 | ## Reproducibility |
| 53 | |
| 54 | The files were generated with audio.cpp's `audiocpp_gguf` converter: |
| 55 | |
| 56 | ```bash |
| 57 | audiocpp_gguf \ |
| 58 | --input model.safetensors \ |
| 59 | --root /path/to/Fun-ASR-Nano-2512-hf \ |
| 60 | --output fun-asr-nano-2512-q8_0.gguf \ |
| 61 | --type q8_0 \ |
| 62 | --family fun_asr_nano \ |
| 63 | --model-spec model_specs/fun_asr_nano.json |
| 64 | ``` |
| 65 | |
| 66 | Both formats were checked with `audiocpp_gguf --inspect` and full reference |
| 67 | audio transcription on CPU and NVIDIA H100 CUDA. |
| 68 | |
| 69 | ## License |
| 70 | |
| 71 | The original model and these converted weights are governed by the |
| 72 | FunASR Model Open Source License Agreement v1.1 distributed with the source |
| 73 | model. Review that agreement before using or redistributing the files. |
| 74 | |