fix(asr/nemotron): native-Swift mel front-end to fix iPadOS cold-start zero output (#739)

[Alex-Wengg]

Summary

Fixes #739 — StreamingNemotronAsrManager produces zero output on iPadOS cold start (empty transcript, works on macOS).

Root cause (verified, and it differs from the issue's stated theory): the ios17.slice_by_index: zero shape error / "Skipped adding default_function to entry point: main" warning comes from the CoreML preprocessor model's flexible RangeDim audio input, not the encoder. CoreML builds the ANE default_function against the range's 1-sample lower bound → mel extraction slices to zero frames → the preprocessor's ANE main entry point is never built. (Confirmed by loading each model on CPU_AND_NE: only the preprocessor warns; the encoder graph is clean. The issue's proposed encoder cache_len re-trace was a misdiagnosis — cache_len only feeds boolean masks, not a data-dependent slice.)

Fix: remove the CoreML preprocessor from both Nemotron streaming managers and compute log-mel natively in Swift (NemotronMelExtractor). No model re-conversion / HF re-upload needed.

Changes

• NemotronMelExtractor — wraps the shared AudioMelSpectrogram, reproducing NeMo's AudioToMelSpectrogramPreprocessor exactly (n_fft=512, win=400, hop=160, 128 mels, symmetric Hann, preemph=0.97, normalize: NA → raw log-mel, no per-feature normalization — the one difference from UnifiedMelExtractor).
• English (StreamingNemotronAsrManager + +Pipeline): swap preprocessor MLModel for the extractor.
• Multilingual (StreamingNemotronMultilingualAsrManager + +Pipeline/+Buffers/+Shared): same swap, with two extractor instances (on-actor + concurrent triple-stage prefetch) so the extractor's non-thread-safe FFT buffers are never shared across the prefetch boundary; preprocessor dropped from SharedNemotronMultilingualModels.
• Fail loudly: load-time encoder probe throws ASRError.encoderInstantiationFailed if the encoder returns all-zero output (instead of silently emitting empty transcripts).

Verification

• Python parity: CoreML preprocessor == NeMo PyTorch raw log-mel (max |Δ| ≈ 9e-3); per-bin means ≈ −15 confirm it is not per-feature normalized.
• English LibriSpeech test-clean (full, 2620 files): WER 560ms 2.71% · 1120ms 2.58% · 2240ms 2.64%.
• Multilingual FLEURS (7 langs, full splits): AVG WER/CER 8.62 / 8.33 / 8.30% at 560/1120/2240ms; meets-or-beats NVIDIA's published [56,0] numbers on all published languages.
• swift build passes; new unit tests for the extractor (shape/determinism/not-normalized) and the new error case. (Note: swift test needs full Xcode for XCTest — runs in CI; built/validated locally + via CLI benchmarks.)

Docs

• New Sources/FluidAudio/ASR/Parakeet/Streaming/Nemotron/benchmark.md (English).
• Refreshed Documentation/ASR/NemotronMultilingual.md to current tiers (560/1120/2240; dropped obsolete 320ms) with freshly measured numbers.

Not covered

• On-device iPad cold-start ANEProgramProcessRequestDirect status=0x12 confirmation requires a physical iPad — not verifiable from a Mac. The mel-parity root-cause case (removing the RangeDim preprocessor) is solid; the probe guarantees the failure can no longer be silent.

[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	8.7x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 6m 10s • 2026-06-28T21:19:12.613Z}

[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: 0m44s}

_{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]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (161.3 KB)

_{Runtime: 0m5s}

_{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	6.62x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	72.2s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.072s	Average chunk processing time
Max Chunk Time	0.144s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m22s • 06/28/2026, 04:58 PM EST}

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

[github-actions]

VAD Benchmark Results

Performance Comparison

Dataset	Accuracy	Precision	Recall	F1-Score	RTFx	Files
MUSAN	92.0%	86.2%	100.0%	92.6%	499.3x faster	50
VOiCES	92.0%	86.2%	100.0%	92.6%	605.6x 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]

ASR Benchmark Results ✅

Status: All benchmarks passed

Parakeet v3 (multilingual)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.57%	0.00%	6.03x	✅
test-other	1.19%	0.00%	3.34x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	5.89x	✅
test-other	1.00%	0.00%	3.69x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.54x	Streaming real-time factor
Avg Chunk Time	1.987s	Average time to process each chunk
Max Chunk Time	3.239s	Maximum chunk processing time
First Token	2.315s	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	0.62x	Streaming real-time factor
Avg Chunk Time	1.447s	Average time to process each chunk
Max Chunk Time	1.874s	Maximum chunk processing time
First Token	1.404s	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: 12m0s • 06/28/2026, 05:08 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]

✅ Nemotron Multilingual Benchmark — FLEURS

FLEURS en_us, chunk 2240ms, 100 samples, B1 fused decode path. Same English audio against both shipped models.

Model	Language	WER	RTFx
latin/ (pruned 2828)	English	7.84%	6.3x
multilingual/ (full 13087)	English	7.91%	6.2x

Logs (tail)

[latin / English]

Language     | Prompt   | WER%   | CER%   | RTFx   | Duration  | Processed | Skipped
--------------------------------------------------------------------------------
en_us        | en-US    | 7.8    | 3.5    | 6.3    | 953.9s    | 100       | -
--------------------------------------------------------------------------------
AVERAGE      | —        | 7.8    | 3.5    | 6.3   


[multilingual / English]

Language     | Prompt   | WER%   | CER%   | RTFx   | Duration  | Processed | Skipped
--------------------------------------------------------------------------------
en_us        | en-US    | 7.9    | 3.6    | 6.2    | 953.9s    | 100       | -
--------------------------------------------------------------------------------
AVERAGE      | —        | 7.9    | 3.6    | 6.2   

[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	9.82x	>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	18.321	17.1	Fetching diarization models
Model Compile	7.852	7.3	CoreML compilation
Audio Load	0.078	0.1	Loading audio file
Segmentation	27.763	26.0	VAD + speech detection
Embedding	106.588	99.8	Speaker embedding extraction
Clustering (VBx)	0.096	0.1	Hungarian algorithm + VBx clustering
Total	106.854	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 • 134.4s processing • Test runtime: 2m 20s • 06/28/2026, 05:15 PM EST}

[github-actions]

✅ Nemotron Multilingual Benchmark — FLEURS

FLEURS en_us, chunk 2240ms, 100 samples, B1 fused decode path. Same English audio against both shipped models.

Model	Language	WER	RTFx
latin/ (pruned 2828)	English	7.84%	4.6x
multilingual/ (full 13087)	English	7.91%	4.7x

Logs (tail)

[latin / English]

Language     | Prompt   | WER%   | CER%   | RTFx   | Duration  | Processed | Skipped
--------------------------------------------------------------------------------
en_us        | en-US    | 7.8    | 3.5    | 4.6    | 953.9s    | 100       | -
--------------------------------------------------------------------------------
AVERAGE      | —        | 7.8    | 3.5    | 4.6   


[multilingual / English]

Language     | Prompt   | WER%   | CER%   | RTFx   | Duration  | Processed | Skipped
--------------------------------------------------------------------------------
en_us        | en-US    | 7.9    | 3.6    | 4.7    | 953.9s    | 100       | -
--------------------------------------------------------------------------------
AVERAGE      | —        | 7.9    | 3.6    | 4.7   

[github-actions]

Speaker Diarization Benchmark Results

Speaker Diarization Performance

Evaluating "who spoke when" detection accuracy

Metric	Value	Target	Status	Description
DER	80.8%	<30%	⚠️	Diarization Error Rate (lower is better)
JER	80.9%	<25%	⚠️	Jaccard Error Rate
RTFx	29.91x	>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	11.208	31.9	Fetching diarization models
Model Compile	4.803	13.7	CoreML compilation
Audio Load	0.058	0.2	Loading audio file
Segmentation	10.522	30.0	Detecting speech regions
Embedding	17.536	50.0	Extracting speaker voices
Clustering	7.014	20.0	Grouping same speakers
Total	35.081	100	Full pipeline

Speaker Diarization Research Comparison

Research baselines typically achieve 18-30% DER on standard datasets

Method	DER	Notes
FluidAudio	80.8%	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 • 35.1s diarization time • Test runtime: 2m 9s • 06/28/2026, 05:17 PM EST}