feat(tts): MOSS-TTS-Nano backend (beta) — streaming zero-shot voice cloning, 48 kHz stereo

[Alex-Wengg]

Summary

Beta — beta model conversion, same footing as NeuTTS and Inflect: the Swift API, published artifacts and output quality may change. Marked as such in TtsBackend, MossTtsNanoManager, the CLI help and both docs pages.

Swift backend for MOSS-TTS-Nano (OpenMOSS, Apache-2.0): 0.1B multilingual streaming TTS with zero-shot voice cloning, 20 languages, native 48 kHz stereo output. Weights: FluidInference/moss-tts-nano-coreml; conversion: FluidInference/mobius#92.

Pipeline per 80 ms frame: Prefill → (Frame → CodecStep → Step)*. Sampling (top-k/top-p/temperature/repetition penalty) is inside the Frame graph; the host supplies uniform randoms, so seed gives reproducible runs.

• MossTtsNanoManager (actor): loadVoice(.en2/.zh1), cloneVoice(audioURL:) (fp32 codec encoder, fetched on demand), synthesize, synthesizeStreaming (AsyncThrowingStream of stereo frames), tokenize.
• MossTtsNanoTokenizer: SentencePiece BPE (nmt_nfkc, dummy prefix, byte fallback) — id-exact vs upstream sentencepiece on 8 vectors (en/zh/ja/ru/emoji/whitespace). The PocketTTS unigram tokenizer cannot serve this model; SentencePieceProto now also parses the piece type.
• MossTtsNanoTextChunker / MossTtsNanoPromptBuilder: upstream sentence→clause→token-budget chunking with inter-chunk pauses; prompt rows built from the repo's config.json (template ids + special tokens, nothing hardcoded).
• CLI: --backend moss-tts-nano with --voice, --clone-voice, --save-voice, --voice-file, --seed, --greedy, --tokens-only; stereo WAV writer.
• Docs: Documentation/TTS/MossTtsNano.md, Documentation/Models.md rows.

Depends on #918 (ModelHub root-level required-file download fix — tokenizer.model). The branch is stacked on that fix; the diff shrinks to the backend once #918 merges.

Validation (M5 Pro, macOS 26.7)

• Tokenizer parity 8/8 vectors via --tokens-only.
• Release CLI, en_2 voice, 9–10 s utterances: RTFx 3.1–3.8×, first audio 0.22–0.40 s; Mandarin zh_1 3 s utterance 2.5×.
• Parakeet ASR (tts-asr-verify --score-only) on two English phrases: 0 % WER on the 27-word phrase, one spurious trailing word on the short one (sampled output; PyTorch upstream scores 10.1 % on the same pair).
• Prefill/Step pinned to CPU+GPU (ANECCompile fails on those graphs); Frame/CodecStep honour the requested compute units.
• Unit tests: tokenizer normalization (+ full BPE parity when the model is cached), chunker, prompt builder, include rule. swift test is unavailable locally (no XCTest); relying on CI.

Not in this PR

• Text normalization (upstream uses WeTextProcessing) — numbers/dates should be verbalized first.
• Stateful (MLState) Step variant to avoid the 38 MB KV round trip per frame.
• tts-benchmark wiring for the new backend.

🤖 Generated with Claude Code

https://claude.ai/code/session_018UHoFANi4DcvU6TzyTPnHH

[github-actions]

PocketTTS Smoke Test ✅

Check	Result
Build	✅
Model download	✅
Model load	✅
Synthesis pipeline	✅
Output WAV	✅ (135.0 KB)

_{Runtime: 0m7s}

_{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: 0m23s}

_{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]

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	7.07x	Real-time factor (higher = faster)
Total Audio	470.6s	Total audio duration processed
Total Time	69.7s	Total processing time

Streaming Metrics

Metric	Value	Description
Avg Chunk Time	0.070s	Average chunk processing time
Max Chunk Time	0.139s	Maximum chunk processing time
EOU Detections	0	Total End-of-Utterance detections

_{Test runtime: 1m22s • 09/13/2026, 08:19 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	18.8x	>1.0x	✅
Speakers	4/4	-	-

_{Sortformer High-Latency • ES2004a • Runtime: 2m 54s • 2026-09-14T00:01:51.985Z}

[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	12.95x	>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	19.988	24.7	Fetching diarization models
Model Compile	8.566	10.6	CoreML compilation
Audio Load	0.032	0.0	Loading audio file
Segmentation	21.284	26.3	VAD + speech detection
Embedding	80.823	99.7	Speaker embedding extraction
Clustering (VBx)	0.103	0.1	Hungarian algorithm + VBx clustering
Total	81.046	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 • 102.2s processing • Test runtime: 1m 53s • 09/13/2026, 08:23 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%	641.6x faster	50
VOiCES	94.0%	89.3%	100.0%	94.3%	678.3x 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%	3.64x	✅
test-other	1.19%	0.00%	2.25x	✅

Parakeet v2 (English-optimized)

Dataset	WER Avg	WER Med	RTFx	Status
test-clean	0.80%	0.00%	4.24x	✅
test-other	1.40%	0.00%	2.92x	✅

Streaming (v3)

Metric	Value	Description
WER	0.00%	Word Error Rate in streaming mode
RTFx	0.44x	Streaming real-time factor
Avg Chunk Time	2.211s	Average time to process each chunk
Max Chunk Time	3.203s	Maximum chunk processing time
First Token	2.736s	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	0.59x	Streaming real-time factor
Avg Chunk Time	1.496s	Average time to process each chunk
Max Chunk Time	2.400s	Maximum chunk processing time
First Token	1.613s	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: 9m14s • 09/13/2026, 08:45 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]

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	16.26x	>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	13.762	21.3	Fetching diarization models
Model Compile	5.898	9.1	CoreML compilation
Audio Load	0.065	0.1	Loading audio file
Segmentation	19.350	30.0	Detecting speech regions
Embedding	32.250	50.0	Extracting speaker voices
Clustering	12.900	20.0	Grouping same speakers
Total	64.519	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 • 64.5s diarization time • Test runtime: 3m 51s • 09/13/2026, 08:39 PM EST}