[Alex-Wengg]

Why is this change needed?

Every SenseVoiceEncoderPrecision.fp32 transcription fails. The fp32 encoder (SenseVoiceSmall_fp32.mlmodelc) has one fixed [1, 1800, 560] input, but runEncoder pads to the fp16 encoder's buckets (SenseVoiceConfig.buckets), so Core ML rejects it:

MultiArray shape (1 x 256 x 560) does not match the shape (1 x 1800 x 560) specified in the model description

Found while reviewing #1082, on FLEURS en_us with fluidaudiocli sensevoice-transcribe <wav> --fp32.

What changes?

SenseVoiceManager reads the lengths its encoder's speech input accepts from the model description once, at init:

• the enumerated buckets for fp16 and int8
• 1800 for fp32

It then pads to the smallest one that fits.

Verification (macOS 27, M5 Pro, cached HF models)

• --fp32 now transcribes. It decodes FLEURS en_us_0001 and en_us_0002 correctly; the fp16 encoder returns empty text on those clips (below).
• fp16 and int8 output is byte-identical to main on 24 FLEURS clips (en/de/fr/es, 3 each, both precisions).
• New SenseVoiceEncoderFrameCountsTests, gated on the cached models: fp16 resolves to SenseVoiceConfig.buckets, fp32 to [1800], and fp32 transcribes the real 16.9 s fixture. 11 SenseVoice tests pass.

Not fixed here

The fp16 and int8 encoders output all-NaN ctc_logits on some clips (2 of 6 FLEURS en clips), so the transcript and tags come back empty. It depends on the content, happens on every compute unit, and the fp32 encoder is clean. That needs a fixed fp16 conversion, not a host change.

🤖 Generated with Claude Code

[github-actions]

Speaker Diarization Benchmark Results

Speaker Diarization Performance

Evaluating "who spoke when" detection accuracy

Metric	Value	Target	Status	Description
DER	15.1%	<30%	✅	Diarization Error Rate (lower is better)
JER	24.9%	<25%	✅	Jaccard Error Rate
RTFx	28.48x	>1.0x	✅	Real-Time Factor (higher is faster)

Diarization Pipeline Timing Breakdown

Time spent in each stage of speaker diarization

Stage	Time (s)	%	Description
Model Download	9.359	25.4	Fetching diarization models
Model Compile	4.011	10.9	CoreML compilation
Audio Load	0.021	0.1	Loading audio file
Segmentation	11.051	30.0	Detecting speech regions
Embedding	18.418	50.0	Extracting speaker voices
Clustering	7.367	20.0	Grouping same speakers
Total	36.840	100	Full pipeline

Speaker Diarization Research Comparison

Research baselines typically achieve 18-30% DER on standard datasets

Method	DER	Notes
FluidAudio	15.1%	On-device CoreML
Research baseline	18-30%	Standard dataset performance

Note: RTFx shown above is from GitHub Actions runner. On Apple Silicon with ANE:

• M2 MacBook Air (2022): Runs at 150 RTFx real-time
• Performance scales with Apple Neural Engine capabilities

_{🎯 Speaker Diarization Test • AMI Corpus ES2004a • 1049.0s meeting audio • 36.8s diarization time • Test runtime: 2m 54s • 11/01/2026, 04:08 PM EST}

[github-actions]

Parakeet EOU Benchmark Results ✅

Status: Benchmark passed
Chunk Size: 320ms
Files Tested: 100/100

Performance Metrics

Metric	Value	Description
WER (Avg)	7.03%	Average Word Error Rate
WER (Med)	4.17%	Median Word Error Rate
RTFx	5.47x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	89.1s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.089s	Average chunk processing time
Max Chunk Time	0.178s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m53s • 11/01/2026, 04:25 PM EST}

_{RTFx = Real-Time Factor (higher is better) • Processing includes: Model inference, audio preprocessing, state management, and file I/O}

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (161.3 KB)

_{Runtime: 0m28s}

_{Note: PocketTTS uses CoreML MLState (macOS 15) KV cache + Mimi streaming state. CI VM lacks physical GPU — audio quality and performance may differ from Apple Silicon.}

[github-actions]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	94.0%	89.3%	100.0%	94.3%	708.5x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	681.0x faster	50

Dataset Details

• MUSAN: Music, Speech, and Noise dataset - standard VAD evaluation
• VOiCES: Voices Obscured in Complex Environmental Settings - tests robustness in real-world conditions

✅: Average F1-Score above 70%