fix(tts/pockettts): per-language mimi encoders for non-English voice cloning (#793)

[Alex-Wengg]

Summary

Fixes the residual flaky/garbled non-English voice cloning reported in #793 (still present after #797).

Root cause

Every pocket-tts language pack ships its own mimi codec weights — all 87 mimi state-dict keys differ across languages, including all 43 encoder-side keys. The shared root mimi_encoderv2.mlmodelc was traced from the English model, so live-clone conditioning for non-English packs was built from the wrong codec's latents. The #797 reprojection corrected the projection, not the latents:

• Swift conditioning vs the reference _encode_audio output: ~140% relative error (mean|diff| 0.149 vs signal 0.107), uniform across frames.
• Injecting the Swift conditioning into the PyTorch reference generation reproduced the exact failure signature ("Maximum generation length reached without EOS" → long noise at low temperature, wrong-voice timbre), while PyTorch-native conditioning was 16/16 clean with the reporter's 5 s reference clip.
• The deployed reprojection assets and the encoder trace itself were verified bit-exact/fp16-exact — the false premise was "mimi is shared across languages" (that July check compared versions of one language, not languages).

Fix

Prefer a pack-local per-language encoder v2.1/<lang>/mimi_encoderv3.mlmodelc (that pack's mimi + speaker_proj baked in; published for all 9 non-English packs on FluidInference/pocket-tts-coreml), fetched lazily on first clone. Its output is already in the pack's own conditioning space → no host-side reprojection. The shared root encoder + #797 reprojection remains the fallback (English, stale caches, offline).

• PocketTtsResourceDownloader.ensurePackMimiEncoder — lazy, best-effort fetch; pack downloads for synthesis-only users stay lean (filter test added).
• PocketTtsModelStore — pack-local-first encoder load; cloneProjection() skips reprojection for the pack-local encoder; 6L warning now scoped to the fallback path.
• CLI: --temperature flag on the PocketTTS path (used to reproduce the reporter's matrix).

Validation (reporter's exact 5.15 s English reference, spanish_24l, 8 seeds × temp {0.15, 0.7})

pipeline	intelligible	noise runaways
PyTorch reference (native cloning)	16/16	0
FluidAudio before (shared encoder + reprojection)	7/16	4 (all at temp 0.15)
FluidAudio after (pack-local encoder)	15/16	0

Conditioning error vs PyTorch native: 140% → 0.06% relative. 6-layer spanish cloning goes from always-garbled to 3/4 intelligible (upstream-level — the earlier "6L broken upstream" verdict was an artifact of injecting English-mimi latents). german_24l verified through the real HF lazy-download path end-to-end.

Conversion script: mobius convert_mimi_encoder_lang.py (companion PR).

Fixes #793

🤖 Generated with Claude Code

[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: 0m30s}

_{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]

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	10.29x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	47.4s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.047s	Average chunk processing time
Max Chunk Time	0.095s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 0m55s • 08/19/2026, 03:13 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	10.4%	<20%	✅	Diarization Error Rate (lower is better)
RTFx	9.84x	>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	21.705	20.4	Fetching diarization models
Model Compile	9.302	8.7	CoreML compilation
Audio Load	0.105	0.1	Loading audio file
Segmentation	28.427	26.7	VAD + speech detection
Embedding	106.363	99.7	Speaker embedding extraction
Clustering (VBx)	0.119	0.1	Hungarian algorithm + VBx clustering
Total	106.653	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 • 134.9s processing • Test runtime: 2m 22s • 08/19/2026, 03:16 PM EST}

[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	17.8x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 3m 9s • 2026-08-19T19:09:33.560Z}

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (146.3 KB)

_{Runtime: 0m20s}

_{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%	489.7x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	495.5x 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	27.50x	>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	14.144	37.1	Fetching diarization models
Model Compile	6.062	15.9	CoreML compilation
Audio Load	0.100	0.3	Loading audio file
Segmentation	11.446	30.0	Detecting speech regions
Embedding	19.077	50.0	Extracting speaker voices
Clustering	7.631	20.0	Grouping same speakers
Total	38.165	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 • 38.2s diarization time • Test runtime: 2m 6s • 08/19/2026, 03:14 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%	4.33x	✅
test-other	1.35%	0.00%	2.68x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	5.16x	✅
test-other	1.56%	0.00%	3.22x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.53x	Streaming real-time factor
Avg Chunk Time	1.719s	Average time to process each chunk
Max Chunk Time	1.961s	Maximum chunk processing time
First Token	2.062s	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	0.57x	Streaming real-time factor
Avg Chunk Time	1.589s	Average time to process each chunk
Max Chunk Time	2.348s	Maximum chunk processing time
First Token	1.674s	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: 12m15s • 08/19/2026, 03:18 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}