MOSS-Transcribe-Diarize-0.9B-CoreML-INT8

Core ML conversion of OpenMOSS-Team/MOSS-Transcribe-Diarize, pinned to revision e6d68cdfcddbdad1a7e8454f0cb859cad76e2502. The model produces timestamped, speaker-attributed text in the form [start][Sxx]text[end].

Model

Property	Value
Parameters	908,513,280 total: 312,463,360 audio/VQ + 596,049,920 text decoder
Architecture	24-layer Whisper-medium encoder, 4x VQ adaptor, 28-layer Qwen3 decoder
Quantization	FP16 audio + symmetric linear INT8 block-32 decoder
Format	Compiled Core ML (.mlmodelc), macOS 15 / iOS 18
Bundle size	1.19 GiB
Audio	16 kHz mono; non-overlapping 30-second encoder chunks
Output	Timestamps, anonymous speaker labels, transcription text
Context	The fixed 1024-token MLState cache covers at most about 1.4 minutes of raw audio before prompt/output tokens.

Files

File	Size	Description
added_tokens.json	707 B	Additional token IDs
audio_encoder.mlmodelc/	596.3 MiB	Compiled Whisper encoder and VQ adaptor
chat_template.jinja	4.7 KiB	Prompt and audio-placeholder template
config.json	844 B	Root loader metadata and download-counting query file
decoder/config.json	1.2 KiB	Decoder geometry, state cache, functions, and precision
decoder/decoder.mlmodelc/	605.1 MiB	Compiled stateful Qwen3 decoder and tied embedding
export_config.json	1.4 KiB	Source revision, artifact hashes, and graph contract
generation_config.json	107 B	Greedy generation token defaults
merges.txt	1.6 MiB	Byte-pair merge rules
preprocessor_config.json	315 B	16 kHz Whisper feature-extractor settings
processing_moss_transcribe_diarize.py	10.7 KiB	Upstream processor implementation
processor_config.json	292 B	Audio-token and timestamp-marker settings
source_config.json	2.3 KiB	Pinned upstream model geometry
special_tokens_map.json	613 B	Special token definitions
tokenizer.json	10.9 MiB	Qwen tokenizer vocabulary and rules
tokenizer_config.json	503 B	Tokenizer configuration
validation.json	2.3 KiB	Measured quality, speed, memory, and parity results
vocab.json	2.6 MiB	Byte-pair vocabulary

Performance

Measured with greedy decoding on an Apple M5 Pro with 48 GB unified memory. Word error rate (WER), character error rate (CER), and real-time factor (RTF) are lower when better. Throughput is 1 / RTF, is higher when better, and reports how many seconds of audio are processed per wall-clock second. RTF excludes model loading. RSS is process memory sampled at clip boundaries; OS high-water RSS also includes transient peaks when the operating system reports it.

Slice	Samples	WER	CER	RTF	Throughput	Sampled / OS high-water RSS	Plain parity vs FP32
english	80	8.32	5.38	0.0571	17.5x real-time	2,937 / 3,614 MB	77/80

Multilingual precision check

Slice	Samples	WER	CER	RTF	Throughput	Plain parity vs FP32
german	10	13.37	5.99	0.0507	19.7x real-time	10/10
french	10	4.91	1.99	0.0803	12.5x real-time	10/10
msa	10	26.50	8.31	0.0620	16.1x real-time	8/10

Round trips

• 30-second overlapping AMI window: RTF 0.113; text=exact, speaker_sequence=exact, raw=drift. The source predicted 3 speakers where RTTM contains 4.
• 33.12-second two-chunk German clip: RTF 0.054; text=exact, speaker_sequence=exact, raw=exact.

The multilingual check covers only ten German, ten French, and ten Modern Standard Arabic clips. It is a conversion check, not proof of the upstream model's full 50-language quality. The meeting round trip checks transcript and speaker-sequence parity but is not a diarization error-rate benchmark.

Usage

Python

from pathlib import Path

import coremltools as ct
from huggingface_hub import snapshot_download

bundle = Path(snapshot_download("aufklarer/MOSS-Transcribe-Diarize-0.9B-CoreML-INT8"))
audio = ct.models.CompiledMLModel(str(bundle / "audio_encoder.mlmodelc"))
embedding = ct.models.CompiledMLModel(
    str(bundle / "decoder/decoder.mlmodelc"), function_name="embedding"
)
decoder = ct.models.CompiledMLModel(
    str(bundle / "decoder/decoder.mlmodelc"), function_name="decoder"
)
state = decoder.make_state()

The host computes Whisper features, replaces audio-placeholder embeddings, runs prompt prefill, then performs greedy token decoding while preserving state. The exact tensor names and shapes are recorded in config.json and decoder/config.json.

Command line

hf download aufklarer/MOSS-Transcribe-Diarize-0.9B-CoreML-INT8 --local-dir ./moss-transcribe-diarize

This downloads a complete low-level model bundle. SDK integration is tracked in speech-swift issue #388; until that integration lands, applications must implement the documented host contract around the exported graphs or weights.

Runtime contract

The multifunction decoder exposes embedding and decoder and uses a persistent FP16 MLState K/V cache. It was loaded with ComputeUnit.ALL and FastPrediction. Prompt construction, audio chunking, generation, cache management, and transcript parsing remain host responsibilities. Machine-readable details and measured results are in config.json, export_config.json, and validation.json.

Source

• Upstream model
• Upstream implementation
• Technical report

No training was performed for this conversion. The source weights contain 908,513,280 parameters and are licensed under Apache 2.0.

Limitations

Speaker IDs are anonymous within each inference. The source model can emit malformed or overlapping timestamps and can miss speakers; conversion parity does not correct those behaviors. ANE-only loading failed with Core ML error -14, so the measured runtime allows Core ML accelerator fallback. The validation here does not establish the upstream 90-minute claim for this deployment format.

Links

• speech-swift — Apple SDK
• Docs — Apple setup and CLI docs
• soniqo.audio — website
• blog — blog

License

Apache License 2.0, inherited from the upstream model.