feat: add int8 variant support for Qwen3-ASR

[Alex-Wengg]

Summary

• Add Qwen3AsrVariant enum (.f32, .int8) so users can choose between full-precision (1.75 GB) and int8-quantized (900 MB) Qwen3-ASR models
• Add Repo.qwen3AsrInt8 case following the existing parakeetEou160/320 pattern
• Add --variant f32|int8 flag to qwen3-benchmark and qwen3-transcribe CLI commands
• Update download(), downloadAndLoad(), defaultCacheDirectory() APIs to accept variant parameter (defaults to .f32)

Benchmark Results (LibriSpeech test-clean, 20 files)

Variant	Avg WER	Median RTFx	Decoder Size
f32	0.8%	2.8x	1.1 GB
int8	1.3%	2.5x	571 MB

Int8 gives ~50% RAM savings with negligible quality impact.

Usage

# CLI
fluidaudio qwen3-benchmark --variant int8 --max-files 20
fluidaudio qwen3-transcribe audio.wav --variant int8

# Library API
let models = try await Qwen3AsrModels.downloadAndLoad(variant: .int8)

Test plan

• swift build -c release compiles cleanly
• qwen3-benchmark --variant int8 --max-files 5 runs and produces valid WER/RTFx
• qwen3-benchmark --max-files 5 (default f32) still works
• CI build passes

🤖 Generated with Claude Code

[claude]

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

---

PR Review: int8 variant support for Qwen3-ASR ✅

Todo List

• Create pending review
• Get PR diff information
• Review code changes for quality and best practices
• Add inline feedback comments
• Submit comprehensive review

Code Review Summary

I've completed a thorough review of this PR which adds int8 variant support for Qwen3-ASR models. The implementation is solid and follows established patterns well.

✅ Key Strengths:

• Consistent API design following existing parakeetEou160/320 pattern
• Proper enum implementation with CaseIterable and Sendable conformance
• Backward compatibility maintained with default .f32 parameter values
• Good logging integration showing variant selection
• CLI help text updated appropriately
• Excellent performance results (50% RAM savings, minimal quality impact)

🔍 Minor Suggestions:

1. Naming consistency: Comment says "FP16 weights" but case is named f32
2. Parameter ordering: Consider placing variant after directory parameter
3. Documentation enhancement: CLI help could include specific RAM usage numbers

Technical Implementation ✅:

• Repo enum properly handles both variants with shared remote path logic
• ModelNames.swift correctly groups both variants for required models
• CLI argument parsing is robust with proper error handling
• No breaking changes to existing APIs

The code quality is high and the feature provides significant value for users with memory constraints. This is ready for merge after CI passes.

---

[github-actions]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	92.0%	86.2%	100.0%	92.6%	695.9x faster	50
VOiCES	92.0%	86.2%	100.0%	92.6%	764.5x 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	14.44x	>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	9.102	12.5	Fetching diarization models
Model Compile	3.901	5.4	CoreML compilation
Audio Load	0.122	0.2	Loading audio file
Segmentation	21.797	30.0	Detecting speech regions
Embedding	36.328	50.0	Extracting speaker voices
Clustering	14.531	20.0	Grouping same speakers
Total	72.695	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 • 72.7s diarization time • Test runtime: 2m 40s • 02/15/2026, 12:12 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	14.5%	<20%	✅	Diarization Error Rate (lower is better)
RTFx	4.46x	>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	10.888	4.6	Fetching diarization models
Model Compile	4.666	2.0	CoreML compilation
Audio Load	0.071	0.0	Loading audio file
Segmentation	26.970	11.5	VAD + speech detection
Embedding	232.006	98.6	Speaker embedding extraction
Clustering (VBx)	2.738	1.2	Hungarian algorithm + VBx clustering
Total	235.272	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 • 261.7s processing • Test runtime: 5m 52s • 02/15/2026, 12:22 AM EST}

[claude]

This PR adds solid int8 variant support for Qwen3-ASR models with excellent API design consistency. The implementation follows established patterns and maintains backward compatibility while providing meaningful performance benefits.

The benchmark results are compelling - achieving 50% RAM reduction (1.75GB → 900MB) with only 0.5% WER increase is an excellent trade-off for resource-constrained scenarios.

Key Strengths:

• ✅ Consistent API design following existing parakeetEou patterns
• ✅ Proper enum with Sendable conformance
• ✅ Backward compatibility maintained with sensible defaults
• ✅ Comprehensive CLI integration with helpful error handling
• ✅ No breaking changes to existing APIs

Minor Suggestions:

1. Consider clarifying the f32 case comment (says "FP16 weights" but case is named f32)
2. CLI help could include specific RAM usage numbers for clarity
3. Consider parameter ordering consistency in downloadAndLoad()

Overall, this is a well-implemented feature that adds valuable functionality without disrupting existing workflows. The code quality is high and follows project conventions.

[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	11.51x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	41.0s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.041s	Average chunk processing time
Max Chunk Time	0.082s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 0m47s • 02/15/2026, 12:12 AM EST}

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

[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%	5.29x	✅
test-other	1.35%	0.00%	3.71x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	5.39x	✅
test-other	1.16%	0.00%	3.52x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.65x	Streaming real-time factor
Avg Chunk Time	1.381s	Average time to process each chunk
Max Chunk Time	1.503s	Maximum chunk processing time
First Token	1.640s	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	0.65x	Streaming real-time factor
Avg Chunk Time	1.394s	Average time to process each chunk
Max Chunk Time	1.577s	Maximum chunk processing time
First Token	1.397s	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: 5m10s • 02/15/2026, 12:13 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]

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	0.1%	-	-
Speaker Error	8.9%	-	-
RTFx	20.8x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 1m 38s • 2026-02-15T05:14:18.222Z}

[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: 5m17s}

_{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.}