feat(asr/v3): opt-in int8-linear Encoder_v2 encoder precision (#760)

[Alex-Wengg]

Root cause (#760)

The shipped v3 Encoder.mlmodelc (labeled "int8", actually Mixed Float16 + 6-bit LUT palettization) corrupts tokens under specific right-context. On the reporter's Ukrainian repro, a ≤15s single-window cut deterministically decodes фоновий процес → non-word фоновиц (confidence 0.885 — silently wrong). Not a window-composition/seam issue: all three cuts decode in a single window.

Full encoder-substitution sweep on identical audio with identical stock companions (Swift CLI):

Encoder	Size	A (tail 2s)	B (tail 6s)	C control
stock 6-bit LUT	425M	✅	❌ фоновиц	✅
int4 (EncoderInt4)	285M	❌ фіновий + deleted clause	⚠️	❌ афоновий
int8 linear per-channel (new)	568M	✅	✅	✅
fp16	1.1G	✅	✅	✅
fp32	2.2G	✅	✅	✅
parakeet-mlx (reference)	—	✅	✅	✅

Each quantization level has its own error sites — int4 passing B in isolation was rounding luck.

Change (strictly opt-in — defaults unchanged)

The re-quantized encoder (int8 per-channel linear, data-free cto.linear_quantize_weights on the fp16 export, +143MB) is published as Encoder_v2.mlmodelc alongside the untouched original (merged on HF).

• .int8 (default) continues to mean the original Encoder.mlmodelc everywhere, including fresh installs. No behavior change for any existing caller.
• New ParakeetEncoderPrecision.int8V2 ("int8-v2") selects Encoder_v2.mlmodelc, via the API parameter or --encoder-precision int8-v2. It rides the existing generic encoderPrecision plumbing — AsrModels.swift is untouched; the diff is the enum case, a filename constant, CLI help text, and tests.
• config.json on HF now records the true weight format of each encoder file (the original is palettized_lut6_mixed_fp16, not int8), so the historical mislabel can't recur.

Verification

• E2E (built CLI, reporter's cuts, model dir containing both encoders): default loads the original (B reproduces фоновиц, proving no silent preference change); --encoder-precision int8-v2 loads v2 and B decodes фоновий процес ✅. Fresh-download of the v2 file from live HF verified.
• Unit tests: filename mapping, variant round-trip, per-precision required-model sets, modelsExist distinguishing the two precisions.
• swift build -c release clean, swift-format lint clean.

Follow-ups (before recommending int8-v2 broadly)

• WER regression run (FLEURS uk + standard ASR benchmark) for the v2 encoder
• ANE RTFx spot-check (linear-quantized weights take a different ANE path than palettized LUTs)

Closes #760

🤖 Generated with Claude Code

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (153.8 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]

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: 0m27s}

_{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]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	94.0%	89.3%	100.0%	94.3%	408.8x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	414.4x 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]

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	8.48x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	58.0s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.058s	Average chunk processing time
Max Chunk Time	0.116s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m40s • 08/19/2026, 04:38 PM EST}

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

[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	16.0x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 3m 27s • 2026-08-19T20:50:10.778Z}

[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	9.22x	>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	22.112	19.4	Fetching diarization models
Model Compile	9.477	8.3	CoreML compilation
Audio Load	0.148	0.1	Loading audio file
Segmentation	29.809	26.2	VAD + speech detection
Embedding	113.578	99.8	Speaker embedding extraction
Clustering (VBx)	0.111	0.1	Hungarian algorithm + VBx clustering
Total	113.828	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 • 143.5s processing • Test runtime: 2m 30s • 08/19/2026, 04:41 PM EST}

[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	16.17x	>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	14.031	21.6	Fetching diarization models
Model Compile	6.013	9.3	CoreML compilation
Audio Load	0.105	0.2	Loading audio file
Segmentation	19.455	30.0	Detecting speech regions
Embedding	32.425	50.0	Extracting speaker voices
Clustering	12.970	20.0	Grouping same speakers
Total	64.891	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 • 64.9s diarization time • Test runtime: 3m 55s • 08/19/2026, 04:48 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%	4.14x	✅
test-other	1.59%	0.00%	2.34x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	4.89x	✅
test-other	1.00%	0.00%	3.24x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.48x	Streaming real-time factor
Avg Chunk Time	1.852s	Average time to process each chunk
Max Chunk Time	2.504s	Maximum chunk processing time
First Token	2.135s	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	1.584s	Average time to process each chunk
Max Chunk Time	1.984s	Maximum chunk processing time
First Token	1.611s	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: 12m48s • 08/19/2026, 04:58 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}