--- library_name: mlx-audio pipeline_tag: automatic-speech-recognition license: apache-2.0 base_model: Reson8/Rumble-2-Flash base_model_relation: quantized language: [en, de, fr, es, it, pt, nl, ru, zh, ja, ko, ar, hi, tr, pl, sv, da, fi, cs, el, ro, hu, id, vi, th, ms, fa, mk] tags: [mlx, speech-to-text, qwen3-asr] --- # Rumble-2-Flash (MLX, 8-bit) [Reson8 Rumble-2-Flash](https://huggingface.co/Reson8/Rumble-2-Flash) for Apple Silicon. Rumble-2-Flash is a 0.5B speech recognition model in the Qwen3-ASR architecture, with a 14-layer text decoder and a 99,630-token vocabulary. The text decoder is quantized to 8 bits (affine, group size 64). The audio encoder and the projector stay in BF16. The weights take 692 MB. ## Conversion Converted from the BF16 checkpoint at revision `a0c34287229103b706c670d70791683e34379732` with [`convert-rumble-2-flash.py`](https://github.com/TypeWhisper/typewhisper-benchmark/blob/main/bench/scripts/convert-rumble-2-flash.py) and mlx-audio 0.6.3: - The transformers 5 weight names are mapped to the original Qwen3-ASR layout that mlx-audio loads: `model.audio_tower.*` to `audio_tower.*`, `model.multi_modal_projector.linear_1/2` to `audio_tower.proj1/proj2` and `model.language_model.*` to `model.*`. The convolution kernels move to MLX's channel-last order. - `config.json` also carries `text_config.rope_theta`, which transformers 5 keeps in `rope_parameters`. - Tokenizer, chat template and processor files are copied unchanged. The optional CTC head of the original repository is left out. ## Usage Rumble-2-Flash numbers `<|im_end|>` and `<|endoftext|>` differently from Qwen3-ASR (99564 and 99562). mlx-audio 0.6.3 reads them from the tokenizer. mlx-audio-swift needs [a fix](https://github.com/TypeWhisper/mlx-audio-swift/tree/seofood/qwen3-asr-tokenizer-eos-on-0.31) for this; without it, generation does not stop at the end of the transcript. The model was trained with three control tokens at the start of the system prompt. Its chat template's defaults are ``: numbers as digits, no filler words, repeated words kept. ```python from mlx_audio.stt import load model = load("TypeWhisper/Rumble-2-Flash-8bit") result = model.generate( "audio.wav", language="German", # leave out to detect the language system_prompt="", ) print(result.text) ``` ## Accuracy Word error rate on the [TypeWhisper benchmark](https://github.com/TypeWhisper/typewhisper-benchmark), mean over the suites both models ran (22 suites in English, German, French, Spanish, Italian, Portuguese and Russian), measured with mlx-audio on a Mac mini: | Model | WER | Real-time factor | |---|---:|---:| | Rumble-2-Flash 8-bit (this repository) | 6.3 % | 0.013 | | Qwen3-ASR 0.6B 8-bit | 10.2 % | 0.035 | The converted model matches Reson8's hosted Rumble-2-Flash API on the 25 suites both were run on. This checkpoint is a model choice in the Qwen3 ASR plugin of [TypeWhisper](https://www.typewhisper.com). ## License Rumble-2-Flash is licensed under [Apache 2.0](https://huggingface.co/Reson8/Rumble-2-Flash/blob/main/LICENSE) by Reson8. The license is included in this repository.