Muesli-HQ/muesli

when I transcribe a meeting, I copy the transcript and provide it to my copilot manually, adding persona and skills as I see fit. Unfortunately, using the "copy" button for the transcription does not include the meeting date. It would be really cool to add all the meeting meta data to the copy along the transcript, so that my copilot will have all the information. thanks!

I'm on a large meeting, mostly muted in Teams. Near me, my colleagues are speaking. ==> Muesli catches my colleagues and transcribe them, but they are not part of the meeting. it would be cool if muesly can understand if I'm muted in teams, and ignore my mic completely.

Summary Explore porting microsoft/VibeVoice-ASR to Core ML and adding it to Muesli as another local transcription runtime. Why this is interesting VibeVoice-ASR is a Microsoft Research ASR model with several properties that could map well to Muesli's meeting transcription use case: Long-form ASR: model card says it can process up to 60 minutes of audio in a single pass. Rich transcript output: designed to emit speaker, timestamp, and content structure together. Multilingual/code-switching support: over 50 languages without requiring explicit language selection. Custom hotwords: supports user-provided terms, names, and domain vocabulary. MIT license. Model card details to account for: Hugging Face model: microsoft/VibeVoice-ASR Architecture/library: Transformers / VibeVoice Model size: about 9B parameters Tensor type: BF16 Paper: arXiv 2601.18184 Proposed implementation path Investigate whether VibeVoice-ASR can be converted to Core ML with a practical Apple Silicon runtime profile

Problem When a meeting is being transcribed, the app currently appears to route dictation and meeting transcription through the same speech-to-text path. In practice this means hotkey dictation is effectively unavailable while a meeting is active. This creates a bad UX for users who want to keep using normal dictation during a meeting, for example to jot down notes, send a message, or capture a private thought without it becoming part of the meeting transcription flow. Desired Outcome Dictation and meeting transcription should be separated so that: meeting transcription can continue uninterrupted while a meeting is active hotkey dictation still works during a meeting dictation output is treated as a separate user action, not merged into the meeting transcript pipeline the UI/state model makes it clear when a meeting is recording versus when ad hoc dictation is running Things To Investigate Which parts of the current codepath are shared between hold-to-talk dictation and meeting tran

Problem Qwen3 ASR is integrated and working in Muesli, but the autoregressive decoder is significantly slower than our CTC-based Parakeet model: Model Latency Architecture Parakeet TDT v3 ~0.13s CTC (parallel) Qwen3 ASR (int8) ~2-3s Autoregressive (sequential) The bottleneck is the CoreML decoder: it generates one token at a time, each requiring a full CoreML inference call. A 10-word transcription = ~15 sequential decoder calls. Proposed Optimization Re-export the Qwen3 decoder CoreML model with batch-4 token input shape, enabling speculative or batched decoding to reduce CoreML calls by up to 4x. Current decoder input shape hidden_states: [1, 1, 1024] (single token) position_cos: [1, 1, 128] position_sin: [1, 1, 128] mask: [1, 1, 1, cache_len] Target decoder input shape hidden_states: [1, 4, 1024] (4 tokens per call) position_cos: [1, 4, 128] position_sin: [1, 4, 128] mask: [1, 1, 4, cache_len] What's Needed PyTorch → CoreML re-export using