MOSS Transcribe Diarize 0.9B

This first-party Soniqo page documents MOSS Transcribe Diarize 0.9B from the local speech-swift / speech-core implementation. Hugging Face bundles are linked below after the integration notes.

Internal Page First

Landing cards and docs menus now point here first; source model and bundle links remain available from this page.

At a Glance

Model	MOSS Transcribe Diarize 0.9B
Role	Batch transcription with model-generated speaker labels and timestamps
Backend	CoreML with a native Accelerate Whisper frontend
Output	Plain transcript plus timestamped speaker segments
Languages	Multilingual MOSS transcription model
License	Verify the source-model and bundle terms for the intended use
Status	Ready through speech transcribe --engine moss and the MossTranscribe Swift product
Source	OpenMOSS-Team/MOSS-Transcribe-Diarize
Swift product	`MossTranscribe`
CLI / runtime	`speech transcribe --engine moss`

Use

The snippet below mirrors the current speech-swift API or command exposed by the repo.

# INT8 is the recommended default.
.build/release/speech transcribe recording.wav --engine moss

# FP16 reference.
.build/release/speech transcribe recording.wav --engine moss --model fp16

Model Links

Implementation Notes

Matched 80-clip English benchmark on an M5 Pro: both variants reached 8.16% aggregate WER and 5.90% CER.
Across two order-reversed runs, INT8 pooled to RTF 0.070 (14.3× real-time) versus FP16 at 0.079 (12.6×), with 2.38–2.40 GB versus 3.69–3.74 GB sampled peak RSS.
The Swift runtime owns exact Whisper log-mel preprocessing, MOSS time-marker prompt construction, audio-embedding injection, MLState cache updates, greedy decoding, and structured output parsing.
INT8 keeps the audio encoder and decoder I/O in FP16 while applying symmetric linear INT8 block-32 quantization to the decoder.
The decoder has a fixed 1024-token combined prompt/output context. The current runtime is batch-only; --stream is rejected.