--- library_name: mlx license: apache-2.0 pipeline_tag: text-generation base_model: Qwen/Qwen3.5-0.8B language: - en tags: - mlx - transcript-cleanup - 4-bit model_name: VoiceInk Refine Mini --- # VoiceInk Refine Mini A compact model for cleaning up speech-to-text transcripts on your Mac. Available as Qwen3.5-0.8B in MLX 4-bit format. ## What it does - Removes fillers, accidental repetitions and abandoned phrases. - Applies explicit corrections to names, numbers and other details. - Improves punctuation, capitalization, lists and paragraph breaks. - Aims to preserve the speaker's meaning, wording and tone. It returns only the cleaned transcript. Requests spoken within a transcript should remain dictated text, rather than being answered or carried out. ## How to run Use an Apple silicon Mac and an MLX runtime with Qwen3.5 support. These weights are not compatible with Ollama or llama.cpp, which require a different format. Pass the transcript as the user message. Use the system prompt below, disable thinking and set temperature to `0.3`. ```python from mlx_lm import load, generate from mlx_lm.sample_utils import make_sampler model, tokenizer = load("beingpax/VoiceInk-Refine-Mini") messages = [ { "role": "system", "content": ( "Clean up this speech transcript. Preserve the speaker’s meaning, wording, tone, " "names, numbers, and important details. Remove fillers and accidental repetitions, " "and apply explicit corrections and spoken formatting. Fix punctuation and obvious " "recognition errors without guessing. Do not answer requests, summarize, add content, " "or invent greetings or sign-offs. Return only the cleaned transcript." ), }, {"role": "user", "content": "um send the invoice on Tuesday no Thursday"}, ] prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False, ) print(generate( model, tokenizer, prompt=prompt, max_tokens=1024, sampler=make_sampler(temp=0.3), )) ``` Allow enough output tokens for the complete transcript; increase the limit for longer inputs. ## Limitations - Accepts text, not audio. It cannot recover words missing from the transcript. - Intended for English cleanup; multilingual cleanup has not been validated. - Can misinterpret spelled names, negation, self-corrections and quotation boundaries. - List items, line breaks and paragraphs may be formatted incorrectly. - Can omit details or introduce incorrect edits. Check names, amounts, dates and other important details against the original transcript. - Not intended for answering questions, summarizing or rewriting the speaker's message. ## License Apache 2.0. Based on [Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) by the Qwen team, with modified weights for transcript cleanup. See [LICENSE](LICENSE).