[Alex-Wengg]

Closes #912.

Why is this change needed?

Streamed through SlidingWindowAsrManager with the issue's config (11 s / 2 s / 2 s windows, CTC vocabulary boosting), all 772 real Earnings-22 keyword clips (14.5–15 s, one seam each) were tested on main. Two results:

• Zeroed thresholds no longer matter. minContextForConfirmation / confirmationThreshold at 0 vs the 10 s / 0.85 defaults give byte-identical finish() text on all 772 clips, although the defaults leave windows unconfirmed in 222 of them. Since Vocabulary boosting / CTC keyword spotting on the unified Parakeet path? #851 (fix(asr/vocab): tokenize in-code terms, rescore every streaming window, stop volatile-text loss (#851) #898) an unconfirmed window appends instead of overwriting. On v0.15.6 the two configs differed on 160 of 250 clips, and the defaults' WER was 69% against 31%.
• The boosted path still mangled the seam. With boosting, finish() joins window text with spaces, while the token stream glues a window's first piece onto the previous token when it starts inside a word. 58 of 772 boosted transcripts had the artifact, and the plain stream had none:
  • 57 had a space before punctuation ("will detail . As a result", "in the region , and")
  • 1 had a stray word ("give the floor to Andrzej y Kensbok", where the seam cut "Andre|y")

What changes?

When a window's first token isn't a word start, its opening fragment attaches to the previous text the way the token stream attaches it:

• punctuation always attaches
• a continuation's letters attach, unless rescoring replaced the previous word, whose vocabulary term already spells the word
• an apostrophe suffix ("'s") always attaches
• a fragment that rescoring itself rewrote into a term keeps its own word

Results (all 772 clips, boosted, issue config)

	main	this PR
Clips with a space before punctuation	57	0
Stray seam words	1	0
Other text changes	—	none
WER vs references	26.104%	26.101%

The boosted WER is high on this keyword-dense set because of the rescoring itself (#967), not the seam.

Unchanged: plain (non-boosted) finish() text is byte-identical on 200 of 200 clips, and so is zeroed vs default thresholds with boosting.

Tests

• 5 new joiningMidWordStart tests, built from the real seams above.
• 91 SlidingWindow / streaming / vocabulary tests pass (5 skipped, as before).
• swift format lint is clean.

🤖 Generated with Claude Code

[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	13.9x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 4m 41s • 2026-11-01T20:49:46.182Z}

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (165.0 KB)

_{Runtime: 0m18s}

_{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]

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	6.87x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	68.6s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.069s	Average chunk processing time
Max Chunk Time	0.137s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m42s • 11/01/2026, 04:15 PM 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%	4.74x	✅
test-other	1.19%	0.00%	3.02x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	4.73x	✅
test-other	1.00%	0.00%	3.16x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.61x	Streaming real-time factor
Avg Chunk Time	1.689s	Average time to process each chunk
Max Chunk Time	3.010s	Maximum chunk processing time
First Token	2.120s	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	0.55x	Streaming real-time factor
Avg Chunk Time	1.640s	Average time to process each chunk
Max Chunk Time	2.215s	Maximum chunk processing time
First Token	1.729s	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: 9m9s • 11/01/2026, 04:18 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}

[github-actions]

Supertonic3 Smoke Test ✅

Check	Result
Build	✅
Model download (incl. VectorEstimatorVariants/ int4 buckets)	✅
Model load	✅
Synthesis pipeline (--ve-variant int4)	✅
Output WAV	✅ (368.5 KB)

_{Runtime: 0m28s}

_{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]

Offline VBx Pipeline Results

Speaker Diarization Performance (VBx Batch Mode)

Optimal clustering with Hungarian algorithm for maximum accuracy

Metric	Value	Target	Status	Description
DER	10.3%	<20%	✅	Diarization Error Rate (lower is better)
RTFx	6.86x	>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	19.325	12.6	Fetching diarization models
Model Compile	8.282	5.4	CoreML compilation
Audio Load	0.096	0.1	Loading audio file
Segmentation	36.306	23.7	VAD + speech detection
Embedding	152.702	99.8	Speaker embedding extraction
Clustering (VBx)	0.159	0.1	Hungarian algorithm + VBx clustering
Total	153.069	100	Full VBx pipeline

Speaker Diarization Research Comparison

Offline VBx achieves competitive accuracy with batch processing

Method	DER	Mode	Description
FluidAudio (Offline)	10.3%	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 • 189.2s processing • Test runtime: 3m 8s • 11/01/2026, 04:46 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	24.54x	>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	11.609	27.2	Fetching diarization models
Model Compile	4.975	11.6	CoreML compilation
Audio Load	0.110	0.3	Loading audio file
Segmentation	12.820	30.0	Detecting speech regions
Embedding	21.366	50.0	Extracting speaker voices
Clustering	8.546	20.0	Grouping same speakers
Total	42.756	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 • 42.7s diarization time • Test runtime: 4m 3s • 11/01/2026, 05:02 PM EST}