mlx-community/VibeVoice-ASR-Streaming-7B-bf16

MLX conversion of microsoft/VibeVoice-ASR-Streaming-7B — streaming speaker-attributed ASR: it transcribes who said what as the audio arrives, with no separate diarization stage. 10 languages, hotword biasing via a context prefix.

Precision: bfloat16 — 16.66 GB. The reference tier for parity work; the 4-bit tier is the deployable one.

What differs from the upstream checkpoint

• The acoustic tokenizer's DECODER is dropped (276 tensors / 0.69 GB). It ships in the upstream repo but ASR never calls it — only the two encoders are used.
• Weights are in the MLX layout (sanitized keys, Conv1d transposed to [O, K, I]).
• Tokenizer and preprocessor_config.json are included and required. The streaming protocol ends each chunk on <|text_chunk_end|> and reads chunk_frames / lookahead_frames from the preprocessor config; a consumer that falls back to the stock Qwen2.5 tokenizer has no such token and never terminates a chunk.

Streaming geometry

speech_tok_compress_ratio 3200 @ 24 kHz = 7.5 frames/s. chunk_frames 22 → 2.9333 s; lookahead_frames 4 → 0.5333 s. So the latency floor is ~3 s: this is live-captioning latency, not dictation latency.

Each chunk is encoded independently from raw samples with cold conv state (split_then_encode) — there is no cross-chunk encoder state to carry.

Measured

On 40.4 min of balanced 4-speaker AMI-IHM (7475 reference words), Apple M5 Max, against the PyTorch reference on identical audio:

tier	cpWER	WER	speakers found	resident	peak	RTF
PyTorch reference	31.76 %	21.25 %	15 / 16	—	—	0.696
MLX bf16	31.68 %	21.17 %	15 / 16	16.66 GB	21.45 GB	0.916
MLX 4-bit	31.95 %	21.77 %	15 / 16	5.72 GB	7.63 GB	0.149

cpWER is concatenated minimum-permutation WER — it charges transcription and speaker attribution. 4-bit costs +0.27 cpWER against bf16 and loses no speakers; that is inside the PyTorch reference's own ±0.8-point run-to-run band (see below), so the degradation sits below the noise floor of the measurement.

Harness: xocialize/mlxengine-audio-tools.

Two things worth knowing

The upstream reference is not deterministic. Its encode_speech samples the acoustic VAE (sample(dist_type='gaussian'), fix_std 0.5) at ~26 % noise-to-signal, so it does not reproduce itself run to run — four runs on identical audio span 1.58 cpWER points. This MLX port takes the distribution's mean instead, which lands within 0.08 cpWER of the stochastic reference and is repeatable.

Quantisation here is device-dependent. CPU- and GPU-computed 4-bit scales disagree on 396 of 1297 tensors. These weights were quantised on GPU (recorded as quantization.mlx_quantised_on in config.json), which is the configuration the numbers above were measured on.

Licence

MIT, from the upstream checkpoint. Conversion by xocialize.