feat: add Qwen3-TTS backend for multilingual text-to-speech

[Alex-Wengg]

Summary

• Add CoreML-based Qwen3-TTS inference pipeline with full prefill → LM decode → code predictor → audio decoder flow
• Support English and Chinese synthesis with temperature+top_k sampling for natural speech and proper EOS detection
• Include automatic silence trimming post-processing for clean audio output

New files

• Qwen3TtsSynthesizer.swift — Full inference pipeline: KV-cache prefill, CB0 sampling with EOS masking, CB1-15 code prediction, audio decoding, and silence trimming
• Qwen3TtsModelStore.swift — CoreML model loading for prefill, decode, code predictor, and audio decoder
• Qwen3TtsManager.swift — High-level API for model loading and synthesis
• Qwen3TtsConstants.swift — Model dimensions, special token IDs, and generation parameters

Modified files

• TtsBackend.swift — Add qwen3Tts case
• TTSCommand.swift — CLI support via --backend qwen3 with bilingual test sentences

Validation

• English ASR (Whisper): exact match across PyTorch, CoreML Python, and Swift pipelines
• Chinese ASR: correct transcription with minor phonetic variance expected from stochastic sampling
• Spectral cosine similarity: 0.73–0.92 between Swift and PyTorch reference (expected range for temperature-sampled TTS)

Test plan

• Build the package with swift build
• Run English synthesis: swift run fluidaudio tts --backend qwen3 "Hello world, this is a test of the text to speech system."
• Run Chinese synthesis: swift run fluidaudio tts --backend qwen3 "你好世界，这是一个文字转语音系统的测试。"
• Verify output WAV files contain intelligible speech at natural duration (~3–5s)

🤖 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	25.86x	>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	8.772	21.6	Fetching diarization models
Model Compile	3.759	9.3	CoreML compilation
Audio Load	0.025	0.1	Loading audio file
Segmentation	12.170	30.0	Detecting speech regions
Embedding	20.283	50.0	Extracting speaker voices
Clustering	8.113	20.0	Grouping same speakers
Total	40.576	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 • 40.6s diarization time • Test runtime: 2m 39s • 02/13/2026, 03:53 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	4.21x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	111.6s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.112s	Average chunk processing time
Max Chunk Time	0.223s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 2m19s • 02/13/2026, 03:57 PM EST}

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

[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	2.97x	>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	13.441	3.8	Fetching diarization models
Model Compile	5.761	1.6	CoreML compilation
Audio Load	0.095	0.0	Loading audio file
Segmentation	36.440	10.3	VAD + speech detection
Embedding	350.247	99.0	Speaker embedding extraction
Clustering (VBx)	2.957	0.8	Hungarian algorithm + VBx clustering
Total	353.868	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 • 389.6s processing • Test runtime: 9m 13s • 02/13/2026, 03:57 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%	3.66x	✅
test-other	1.96%	0.00%	3.01x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	4.15x	✅
test-other	1.00%	0.00%	3.20x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.58x	Streaming real-time factor
Avg Chunk Time	1.620s	Average time to process each chunk
Max Chunk Time	2.805s	Maximum chunk processing time
First Token	1.741s	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	0.48x	Streaming real-time factor
Avg Chunk Time	1.817s	Average time to process each chunk
Max Chunk Time	2.448s	Maximum chunk processing time
First Token	1.901s	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: 8m12s • 02/13/2026, 03:50 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]

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	14.6x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 3m 37s • 2026-02-13T20:45:32.459Z}

[github-actions]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	92.0%	86.2%	100.0%	92.6%	287.9x faster	50
VOiCES	92.0%	86.2%	100.0%	92.6%	312.5x 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%