Fix offline diarizer PLDA parameters download

[Alex-Wengg]

Problem

The offline diarizer benchmark was failing in CI with this error:

[ERROR] [FluidAudio.DiarizationBench] ❌ Failed to initialize models: 
processingFailed("PLDA parameters file not found in /Users/runner/Library/Application Support/FluidAudio/Models")

This occurred in the diarization-benchmark.yml workflow when running:

swift run -c release fluidaudiocli diarization-benchmark --mode offline --auto-download --single-file ES2004a

Root Cause

The OfflineDiarizer.requiredModels set only included the 4 .mlmodelc files:

• Segmentation.mlmodelc
• FBank.mlmodelc
• Embedding.mlmodelc
• PldaRho.mlmodelc

But it was missing plda-parameters.json, which is required by OfflineDiarizerModels.loadPLDAPsi() to load the PLDA psi parameters for VBx clustering.

When DownloadUtils.downloadRepo() downloads models, it only downloads files in the requiredModels set. Since plda-parameters.json wasn't in the set, it never got downloaded, causing the initialization to fail.

Solution

Added plda-parameters.json to the OfflineDiarizer.requiredModels set in ModelNames.swift.

Testing

• Build succeeds
• CI workflow should now pass

🤖 Generated with Claude Code

---

[claude]

Claude finished @Alex-Wengg's task —— View job

---

I'll analyze this and get back to you.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 1 additional finding.

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (180.0 KB)

_{Runtime: 0m43s}

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

[github-actions]

Qwen3-ASR int8 Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Transcription pipeline	✅
Decoder size	571 MB (vs 1.1 GB f32)

_{Runtime: 4m8s}

_{Note: CI VM lacks physical GPU — CoreML MLState (macOS 15) KV cache produces degraded results on virtualized runners. On Apple Silicon: ~1.3% WER / 2.5x RTFx.}

[github-actions]

Sortformer High-Latency Benchmark Results

ES2004a Performance (30.4s latency config)

Metric	Value	Target	Status
DER	33.4%	<35%	✅
Miss Rate	24.4%	-	-
False Alarm	0.2%	-	-
Speaker Error	8.8%	-	-
RTFx	14.4x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 2m 28s • 2026-03-28T20:40:05.115Z}

[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	10.67x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	45.6s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.046s	Average chunk processing time
Max Chunk Time	0.091s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m8s • 03/28/2026, 04:44 PM EST}

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

[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	27.32x	>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	9.551	24.9	Fetching diarization models
Model Compile	4.093	10.7	CoreML compilation
Audio Load	0.042	0.1	Loading audio file
Segmentation	11.518	30.0	Detecting speech regions
Embedding	19.197	50.0	Extracting speaker voices
Clustering	7.679	20.0	Grouping same speakers
Total	38.404	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 • 38.4s diarization time • Test runtime: 2m 21s • 03/28/2026, 04:54 PM 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	14.5%	<20%	✅	Diarization Error Rate (lower is better)
RTFx	5.30x	>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	13.908	7.0	Fetching diarization models
Model Compile	5.960	3.0	CoreML compilation
Audio Load	0.051	0.0	Loading audio file
Segmentation	24.087	12.2	VAD + speech detection
Embedding	197.272	99.6	Speaker embedding extraction
Clustering (VBx)	0.741	0.4	Hungarian algorithm + VBx clustering
Total	198.161	100	Full VBx pipeline

Speaker Diarization Research Comparison

Offline VBx achieves competitive accuracy with batch processing

Method	DER	Mode	Description
FluidAudio (Offline)	14.5%	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 • 222.1s processing • Test runtime: 3m 47s • 03/28/2026, 04:54 PM EST}

[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%	5.90x	✅
test-other	1.19%	0.00%	3.75x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	5.95x	✅
test-other	1.00%	0.00%	3.91x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.70x	Streaming real-time factor
Avg Chunk Time	1.301s	Average time to process each chunk
Max Chunk Time	1.366s	Maximum chunk processing time
First Token	1.547s	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	0.69x	Streaming real-time factor
Avg Chunk Time	1.304s	Average time to process each chunk
Max Chunk Time	1.361s	Maximum chunk processing time
First Token	1.305s	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: 5m51s • 03/28/2026, 04:55 PM 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]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	92.0%	86.2%	100.0%	92.6%	767.1x faster	50
VOiCES	92.0%	86.2%	100.0%	92.6%	788.4x 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%