Title: [Feature Request] Add offline file transcription support for Qwen3-ASR-1.7B and MOSS-Transcribe-Diarize

Is your feature request related to a problem? Please describe.

Memo currently does not provide an integrated local file-transcription workflow for either:

1. Qwen3-ASR-1.7B, including an Apple Silicon-optimized MLX version; or
2. MOSS-Transcribe-Diarize, which jointly produces transcription, timestamps, and speaker labels.

Memo v1.8.3 includes the experimental Qwen3-ASR-0.6B-Int8 model under Sherpa ONNX, but Sherpa ONNX is still unavailable in the normal local file-transcription engine selector, as described in #422.

Even if #422 is fixed, Qwen3-ASR-0.6B Sherpa ONNX is a different model and runtime from Qwen3-ASR-1.7B MLX. Users who want the larger model currently cannot select it for normal audio or video file transcription.

For multi-speaker recordings, Memo also lacks a single-pass local backend that directly returns timestamped speaker-attributed transcription. MOSS-Transcribe-Diarize is particularly suitable for meetings, interviews, podcasts, lectures, and other long-form recordings.

Describe the solution you'd like

Please add two selectable local file-transcription backends.

1. Qwen3-ASR-1.7B

Suggested model for Apple Silicon:

https://huggingface.co/aufklarer/Qwen3-ASR-1.7B-MLX-8bit

Requested behavior:

• Allow the model to be downloaded and managed from Memo’s Transcription settings.
• Expose Qwen3-ASR as an engine in the Home screen’s Local Model → Select Engine menu.
• Support imported audio and video files, not only live/real-time transcription.
• Support automatic language identification and multilingual transcription.
• Show model-loading and transcription progress and allow cancellation.
• Convert the output into normal editable Memo transcript segments.
• Preserve timestamps through a supported alignment workflow where available.
• Store the model in Memo’s configurable model directory without requiring command-line setup.

This should be implemented as a dedicated Qwen3/MLX backend rather than treating the MLX Safetensors model as a Sherpa ONNX model.

2. MOSS-Transcribe-Diarize

Upstream project and model:

https://github.com/OpenMOSS/MOSS-Transcribe-Diarize

https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize

Requested behavior:

• Add MOSS-Transcribe-Diarize as a local file-transcription engine.

• Support long-form audio and video transcription.

• Parse the model’s timestamped speaker format, for example:

  [0.48][S01]Welcome everyone[1.66]

• Map [S01], [S02], etc. to Memo speaker labels and speaker-colored transcript segments.

• Preserve start/end timestamps in Memo’s timeline.

• Allow speakers to be renamed after transcription.

• Preserve speaker information when exporting to formats that support it.

• Provide an option to enable or disable speaker diarization.

• Support hotwords or prompt text if the selected runtime exposes them.

• Report unsupported hardware or insufficient memory clearly instead of silently falling back to another engine.

A local runtime would be preferred. An optional OpenAI-compatible custom endpoint could also be useful for users who run MOSS on another machine.

Suggested UI

Under Settings → Transcription:

• Qwen3-ASR
  • Qwen3-ASR-1.7B MLX 8-bit
  • Installed/download status
  • Model directory
• MOSS Transcribe Diarize
  • Installed/download status
  • Local or custom endpoint mode
  • Speaker diarization toggle
  • Hotwords/prompt field

Under the Home screen’s file-transcription controls:

• Local Whisper
• FunASR
• Parakeet
• Qwen3-ASR
• MOSS Transcribe Diarize

Suggested acceptance criteria

• Both engines can be selected for an imported audio or video file.
• Transcription runs without requiring DevTools or manual configuration edits.
• Progress and cancellation are available.
• Results are saved as normal editable Memo projects.
• MOSS speaker identities remain consistent across transcript segments.
• Timestamp and speaker metadata survive project reopening.
• No audio or transcript is sent to a cloud service when local mode is selected.
• Existing Whisper, FunASR, and Parakeet workflows continue to work unchanged.

Describe alternatives you've considered

• Using Whisper or FunASR inside Memo. These do not provide the same Qwen3-ASR-1.7B model quality or MOSS’s joint transcription/diarization output.
• Using BaoCut for transcription and importing an SRT file into Memo. This works for basic subtitles, but it creates a separate workflow and may lose structured speaker metadata.
• Restoring Sherpa ONNX in the local engine selector through #422. This would make Qwen3-ASR-0.6B usable, but would not add Qwen3-ASR-1.7B MLX or MOSS-Transcribe-Diarize.

Additional context

Environment used to confirm the current limitation:

• Memo: v1.8.3
• macOS: 15.8
• Architecture: arm64 / Apple Silicon

Related issues:

• #422 — Sherpa ONNX/Qwen3 is absent from the local file-transcription engine selector
• #305 — Speaker diarization result handling/export

Begin with the Home screen’s Local Model → Select Engine flow and Settings → Transcription, then review related issues #422 and #305 for existing engine-selector and speaker-result context. Done means both local backends work for imported audio or video, with progress, cancellation, editable transcript results, timestamps, and preserved speaker metadata without cloud transmission.