feat(asr/canary): Canary-1B-v2 AED engine + CTC-spotter custom vocab [beta]

[Alex-Wengg]

Note

Beta launch — this backend ships as a beta: the public API surface and hosted model assets may still change in response to feedback before it is declared stable.

Summary

Adds NVIDIA Canary-1B-v2 (attention encoder-decoder, AED) as a selectable on-device ASR engine, converted to CoreML and running int4 on the Neural Engine. Canary is a more accurate transcriber than the existing TDT/CTC paths (esp. on hard domains), and this PR also gives it custom-vocabulary support by reusing the existing CTC keyword spotter.

Users pick the engine that fits: canary (best WER) or the existing ctc custom-vocab path (fastest, top keyword recall).

What's included

• CanaryManager — actor; pipeline: fp32/CPU mel preprocessor → FastConformer encoder → autoregressive transformer decoder + 1024→16384 projection → greedy decode to EOS. Reads the decoder sequence length from the model so a shorter export is picked up automatically.
• CanaryModels — download/load from FluidInference/canary-1b-v2-coreml (int4 ANE default / fp16 parity / int8 CPU); CanaryPrecision.
• CanaryKeywordBooster — custom-vocab support for canary by reusing CtcKeywordSpotter: fuzzy-replace mis-transcribed terms, plus timestamp-guided insertion of keywords canary missed entirely. Precision-protected via a score floor.
• CLI — canary-transcribe (file + LibriSpeech benchmark) and canary-earnings-benchmark (Earnings22-keywords, OpenBench-comparable WER + keyword P/R/F1).
• ModelNames — Repo.canary1bV2, ModelNames.Canary, wired into getRequiredModelNames.
• Tests — CanaryConfigTests (registration / precision / config contract).

Benchmarks (int4, Apple Silicon ANE)

LibriSpeech test-clean (≤15s): WER ~1.7%, RTFx ~10.8x.

Earnings22-keywords (full 772 chunks), scored by the same whole-word keyword metric as Argmax's OpenBench:

Engine	WER	Keyword F1	RTFx
CTC custom-vocab (existing)	22.5%	0.97	35.8x
Canary + vocab + injection	16.5%	0.95	10.8x

Both beat Argmax's published parakeet-v2/v3 keyword F1 (0.91 / 0.89). Canary additionally wins WER by ~6 points.

Model conversion

CoreML conversion pipeline lives in the mobius repo (models/stt/canary-1b-v2/coreml/): NeMo→CoreML export, int4/int8 quantization, projection model, validation (byte-exact vs PyTorch greedy decode), HF staging. Models hosted at FluidInference/canary-1b-v2-coreml.

Notes / follow-ups

• int4 requires iOS18 / macOS15 (int4 weight payloads); fp16 is the iOS17 fallback.
• Decoder has no KV cache yet — re-runs the sequence each step, so canary is ~3x slower than the CTC path. A cache-external decoder export is the planned follow-up to close the gap.
• 15s window per decode; audio >15s is now chunked into overlapping 15s windows (3s overlap) and stitched at the seam via token-level longest-common-substring. Audio ≤15s is unchanged.

🤖 Generated with Claude Code

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (172.5 KB)

_{Runtime: 1m16s}

_{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	11.28x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	42.8s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.043s	Average chunk processing time
Max Chunk Time	0.086s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m35s • 08/18/2026, 08:31 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	13.34x	>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	24.499	31.1	Fetching diarization models
Model Compile	10.500	13.3	CoreML compilation
Audio Load	0.059	0.1	Loading audio file
Segmentation	21.466	27.3	VAD + speech detection
Embedding	78.453	99.7	Speaker embedding extraction
Clustering (VBx)	0.097	0.1	Hungarian algorithm + VBx clustering
Total	78.659	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 • 100.0s processing • Test runtime: 2m 0s • 08/18/2026, 08:39 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%	5.34x	✅
test-other	1.19%	0.00%	2.98x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	3.91x	✅
test-other	1.00%	0.00%	3.43x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.57x	Streaming real-time factor
Avg Chunk Time	1.644s	Average time to process each chunk
Max Chunk Time	2.175s	Maximum chunk processing time
First Token	2.156s	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	0.58x	Streaming real-time factor
Avg Chunk Time	1.512s	Average time to process each chunk
Max Chunk Time	1.664s	Maximum chunk processing time
First Token	1.507s	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: 11m46s • 08/18/2026, 08:50 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]

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: 0m32s}

_{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]

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	17.3x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 3m 6s • 2026-08-19T00:39:21.210Z}

[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	26.83x	>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	12.062	30.8	Fetching diarization models
Model Compile	5.170	13.2	CoreML compilation
Audio Load	0.069	0.2	Loading audio file
Segmentation	11.725	30.0	Detecting speech regions
Embedding	19.542	50.0	Extracting speaker voices
Clustering	7.817	20.0	Grouping same speakers
Total	39.105	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 • 39.1s diarization time • Test runtime: 2m 34s • 08/18/2026, 08:47 PM EST}

[github-actions]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	94.0%	89.3%	100.0%	94.3%	711.0x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	689.1x 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%