[Alex-Wengg]

Refs #1063.

Why is this change needed?

A Parakeet v3 language hint only selects a script. Every Latin-script Language resolves to .latin, so .spanish, .english and the rest filter the same tokens. On FLEURS es_419 (dev 408 + test 100 clips), .spanish, .english and no hint gave byte-identical output on all 508 clips; on 100 clips each of it/pt/de, the native hint matched no hint every time. Only .french differs, because it also runs the English blocklist from #630.

The issue's failure reproduces on read speech: in 6 of 508 Spanish clips v3 emits English function words, and in 2 it renders most of the sentence in English ("…estos plans cayeron in disuso … when most de 80 000 combatientes … and formed the front…").

What changes?

The docs and help text now say what the hint does, so callers don't read it as a per-language constraint:

• the Language doc comment
• the CLI --language help
• a "Scope: script, not language" section in TokenLanguageFilter.md

They point to the two decoders in FluidAudio that are conditioned on a language: Canary (CanaryManager(models:source:target:)) and Cohere Transcribe (language:).

FLEURS es_419	WER dev / test	Clips with English function words
v3 (any hint, or none)	4.86% / 4.03%	6
Canary-1b-v2, source/target .spanish	4.50% / 4.94%	0

Not changed: a Spanish English-blocklist

I tried a Spanish blocklist: the #630 list minus pieces that start a word in the FLEURS es train text, plus the/ The.

• It moved dev WER from 4.86% to 4.70%, and no clip got worse.
• The two translated clips stayed mostly wrong: 51 → 46 and 35 → 26 word errors, against Canary's 0 and 1.
• Its largest gain comes from blocking The, which also hits English titles in Spanish text ("The Guardian", in 11 of 2,796 train sentences).
• Without that piece, it does nothing.
• The fix(asr): reduce English drift on French recordings via token blocklist #630 list as-is turns "hermana" into "armana" and "Washington" into "Witthington".

Validating a list like this needs spontaneous Spanish and code-switched audio first (#840).

🤖 Generated with Claude Code

[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	10.79x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	44.0s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.044s	Average chunk processing time
Max Chunk Time	0.088s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m0s • 11/01/2026, 01:54 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%	5.76x	✅
test-other	1.19%	0.00%	3.30x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	5.79x	✅
test-other	1.00%	0.00%	3.77x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.62x	Streaming real-time factor
Avg Chunk Time	1.446s	Average time to process each chunk
Max Chunk Time	1.598s	Maximum chunk processing time
First Token	1.737s	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	0.66x	Streaming real-time factor
Avg Chunk Time	1.356s	Average time to process each chunk
Max Chunk Time	1.525s	Maximum chunk processing time
First Token	1.351s	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: 6m26s • 11/01/2026, 02:19 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: 0m26s}

_{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]

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	16.58x	>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	14.434	22.8	Fetching diarization models
Model Compile	6.186	9.8	CoreML compilation
Audio Load	0.061	0.1	Loading audio file
Segmentation	18.980	30.0	Detecting speech regions
Embedding	31.634	50.0	Extracting speaker voices
Clustering	12.654	20.0	Grouping same speakers
Total	63.298	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 • 63.3s diarization time • Test runtime: 4m 26s • 11/01/2026, 02:46 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	15.6x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 4m 0s • 2026-11-01T21:03:06.759Z}

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (157.5 KB)

_{Runtime: 0m19s}

_{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.}