feat(benchmark): fetch AMI corpus from HuggingFace mirror, Edinburgh as fallback

[Alex-Wengg]

Follow-up to #754 / #752.

Problem

The AMI annotations and audio are downloaded from groups.inf.ed.ac.uk on every cache-miss CI run. That server is intermittently unavailable — it was hard-down (connection refused) while this change was developed and tested — and a single transient failure previously poisoned a CI benchmark run with placeholder ground truth (#752). #754 made failures loud and cached the annotations in CI; this PR removes the single-point dependency entirely.

Changes

• Mirrored the AMI data used by the benchmarks to FluidInference/ami-corpus-mirror (CC BY 4.0, redistributed with attribution):
  • annotations/ami_public_manual_1.6.2.zip — AMI public manual annotations v1.6.2
  • sdm/{meeting}.Mix-Headset.wav — the official 16-meeting AMI-SDM evaluation split
• DatasetDownloader now tries the mirror first for both the annotations zip and per-meeting audio; the Edinburgh URLs remain as fallback. Meetings/variants absent from the mirror (e.g. IHM) 404 there and fall through to upstream.

Verification

Ran with the Edinburgh server unreachable (verified connection refused immediately before and after the test):

• diarization-benchmark --auto-download --single-file ES2004a --threshold 0.7 fetched the annotations from the mirror (687 segment files extracted) and reproduced the healthy baseline: DER 15.1% / JER 24.8%, exit 0
• swift format lint clean, release build clean

🤖 Generated with Claude Code

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (176.3 KB)

_{Runtime: 0m54s}

_{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]

ASR Benchmark Results ✅

Status: All benchmarks passed

Parakeet v3 (multilingual)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.57%	0.00%	6.66x	✅
test-other	1.35%	0.00%	3.93x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	6.79x	✅
test-other	1.00%	0.00%	3.87x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.70x	Streaming real-time factor
Avg Chunk Time	1.291s	Average time to process each chunk
Max Chunk Time	1.363s	Maximum chunk processing time
First Token	1.536s	Latency to first transcription token
Total Chunks	31	Number of chunks processed

Streaming (v2)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.68x	Streaming real-time factor
Avg Chunk Time	1.298s	Average time to process each chunk
Max Chunk Time	1.413s	Maximum chunk processing time
First Token	1.289s	Latency to first transcription token
Total Chunks	31	Number of chunks processed

_{Streaming tests use 5 files with 0.5s chunks to simulate real-time audio streaming}

_{25 files per dataset • Test runtime: 11m14s • 07/06/2026, 12:12 AM EST}

_{RTFx = Real-Time Factor (higher is better) • Calculated as: Total audio duration ÷ Total processing time
Processing time includes: Model inference on Apple Neural Engine, audio preprocessing, state resets between files, token-to-text conversion, and file I/O
Example: RTFx of 2.0x means 10 seconds of audio processed in 5 seconds (2x faster than real-time)}

Expected RTFx Performance on Physical M1 Hardware:

• M1 Mac: ~28x (clean), ~25x (other)
• CI shows ~0.5-3x due to virtualization limitations

_{Testing methodology follows HuggingFace Open ASR Leaderboard}

[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	12.40x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	38.6s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.039s	Average chunk processing time
Max Chunk Time	0.077s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m48s • 07/06/2026, 12:13 AM EST}

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

[github-actions]

Supertonic3 Smoke Test ✅

Check	Result
Build	✅
Model download (incl. VectorEstimatorVariants/ int4 buckets)	✅
Model load	✅
Synthesis pipeline (--ve-variant int4)	✅
Output WAV	✅ (364.7 KB)

_{Runtime: 0m23s}

_{Note: CI VMs lack a physical Neural Engine; the ANE-bucketed VectorEstimator falls back to CPU here. This validates download + variant resolution + synthesis, not ANE residency/perf.}

[github-actions]

Sortformer High-Latency Benchmark Results

ES2004a Performance (30.4s latency config)

Metric	Value	Target	Status
DER	30.3%	<35%	✅
Miss Rate	28.2%	-	-
False Alarm	0.9%	-	-
Speaker Error	1.2%	-	-
RTFx	22.6x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 2m 56s • 2026-07-06T04:18:01.252Z}

[github-actions]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	92.0%	86.2%	100.0%	92.6%	794.2x faster	50
VOiCES	92.0%	86.2%	100.0%	92.6%	808.4x 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%

[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	31.71x	>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	9.169	27.7	Fetching diarization models
Model Compile	3.929	11.9	CoreML compilation
Audio Load	0.033	0.1	Loading audio file
Segmentation	9.927	30.0	Detecting speech regions
Embedding	16.545	50.0	Extracting speaker voices
Clustering	6.618	20.0	Grouping same speakers
Total	33.097	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 • 33.1s diarization time • Test runtime: 2m 19s • 07/06/2026, 12:20 AM EST}

[github-actions]

Offline VBx Pipeline Results

Speaker Diarization Performance (VBx Batch Mode)

Optimal clustering with Hungarian algorithm for maximum accuracy

Metric	Value	Target	Status	Description
DER	10.4%	<20%	✅	Diarization Error Rate (lower is better)
RTFx	9.37x	>1.0x	✅	Real-Time Factor (higher is faster)

Offline VBx Pipeline Timing Breakdown

Time spent in each stage of batch diarization

Stage	Time (s)	%	Description
Model Download	15.867	14.2	Fetching diarization models
Model Compile	6.800	6.1	CoreML compilation
Audio Load	0.046	0.0	Loading audio file
Segmentation	31.846	28.4	VAD + speech detection
Embedding	111.736	99.7	Speaker embedding extraction
Clustering (VBx)	0.105	0.1	Hungarian algorithm + VBx clustering
Total	112.016	100	Full VBx pipeline

Speaker Diarization Research Comparison

Offline VBx achieves competitive accuracy with batch processing

Method	DER	Mode	Description
FluidAudio (Offline)	10.4%	VBx Batch	On-device CoreML with optimal clustering
FluidAudio (Streaming)	17.7%	Chunk-based	First-occurrence speaker mapping
Research baseline	18-30%	Various	Standard dataset performance

Pipeline Details:

• Mode: Offline VBx with Hungarian algorithm for optimal speaker-to-cluster assignment
• Segmentation: VAD-based voice activity detection
• Embeddings: WeSpeaker-compatible speaker embeddings
• Clustering: PowerSet with VBx refinement
• Accuracy: Higher than streaming due to optimal post-hoc mapping

_{🎯 Offline VBx Test • AMI Corpus ES2004a • 1049.0s meeting audio • 143.7s processing • Test runtime: 2m 23s • 07/06/2026, 12:23 AM EST}