webgpu_browser_validation.md

Browser WebGPU validation

Test configuration

Measured in Chrome 150 on an NVIDIA RTX 5090 with driver 595.84, using Emscripten 6.0.13, Debug builds with -O3 -g, WASM SIMD, --backend webgpu, one model thread and logging enabled. Requests run sequentially.

Warmed rows report the mean of three requests in the same model session after one excluded warmup request. Other rows report a single first request. Timing excludes model loading and WASM initialization. First-request timings include graph/shader setup performed during inference.

RTF is request time divided by generated audio duration, or input duration for ASR, alignment and separation, using canonical session.wall_ms. RTF below 1 means faster than real time. Audio seconds and request seconds are per request, including for warmed rows.

Peak VRAM is the observed NVML graphics-memory peak of a dedicated Chrome GPU process, sampled every 20 ms across loading, warmup and inference. It includes the browser baseline and may miss short peaks.

Results

Model	Task	GGUF format	Input	Audio seconds	Request seconds	RTF	Peak VRAM (MiB)	WASM
Irodori-TTS Small MF	TTS	Q8_0	205 chars	41.800	4.766	0.1140	3581.3	32
Irodori-TTS Small RF	TTS	Q8_0	205 chars	41.800	13.736	0.3286	3708.9	32
Kokoro-82M	TTS	Q8_0	245 chars	15.075	1.086	0.0720	1811.0	32
LFM3-Audio-2B	TTS	Q8_0 + F16 codec	245 chars; warmed	15.360	5.150	0.3353	2146.4	64
Magpie Multilingual	TTS	Q8_0	245 chars	13.375	50.691	3.7901	1457.2	32
Ming3-TTS-0.5B	TTS	Q8_0	245 chars	12.500	20.262	1.6210	1385.1	32
MOSS-TTS Nano	TTS	Q8_0	245 chars	14.000	15.075	1.0768	561.6	32
OmniVoice	TTS	Q8_0	245 chars	15.000	5.027	0.3351	1776.1	32
PocketTTS	TTS	Q8_0	245 chars	11.600	2.062	0.1778	735.4	32
PocketTTS Small 18M	TTS	F32	245 chars; warmed	11.040	3.360	0.3043	260.9	32
Qwen3-TTS 1.7B CustomVoice	TTS	Q8_0	245 chars	18.080	21.590	1.1941	3827.7	64
Supertonic 3	TTS	F32/I64 (Q8-labelled)	245 chars	16.023	0.687	0.0429	502.7	32
VibeVoice 1.5B	TTS	Q8_0	256 chars	13.200	21.878	1.6574	4501.0	64
VoxCPM-0.5B	TTS	Q8_0	245 chars	15.360	15.783	1.0276	2432.9	32
Cohere Transcribe	ASR	Q4_0	40 s audio	40.000	1.548	0.0387	2138.8	32
Fun-ASR Nano	ASR	Q8_0	40 s audio	40.000	2.609	0.0652	1795.4	32
GigaAM Multilingual Large	ASR	F16	40 s audio	40.000	0.329	0.0082	1216.2	32
Moonshine Streaming Medium	ASR	Q8_0	40 s audio	40.000	1.628	0.0407	730.2	32
LFM3-Audio-2B	ASR	Q8_0 + F16 codec	40 s audio; warmed	40.000	1.601	0.0400	2358.9	64
MOSS Transcribe Diarize	ASR	Q4_K	40 s audio	40.000	4.748	0.1187	1642.2	32
Nemotron-3.5-ASR-Streaming-0.6B	ASR	Q8_0	40 s audio	40.000	32.524	0.8131	1111.0	32
OWSM-CTC-v4-1B	ASR	Q8_0	40 s audio	40.000	1.115	0.0279	2891.5	32
Parakeet-TDT-0.6B-v3	ASR	Q8_0	40 s audio	40.000	1.742	0.0436	2079.5	32
Qwen3-ASR-0.6B	ASR	Q8_0	40 s audio	40.000	2.334	0.0584	2359.9	32
SenseVoice-Small	ASR	Q8_0	40 s audio	40.000	0.425	0.0106	369.3	32
Qwen3 forced aligner	ALIGN	Q8_0	40 s audio	40.000	0.365	0.0091	1978.7	32
BS-RoFormer	SEP	Q8_0	40 s audio	40.000	37.927	0.9482	2133.7	32
HTDemucs	SEP	Q8_0	40 s audio	40.000	1.689	0.0422	502.0	32
YuE2	GEN	Q4_0 + F16 VAE	Instrumental; 1,500-token cap	59.999	37.021	0.6170	3014.4	64
NVIDIA NemotronLabs RE-USE	S2S enhancement	F32	14.07 s audio; offline	14.072	4.335	0.3080	3090.0	32
Realtime RE-USE	S2S enhancement	F32	14.07 s audio; streaming	14.072	1.569	0.1115	522.5	32

Workload and configuration

• TTS uses 245 English characters; Irodori uses 205 Japanese characters. VibeVoice's 256-character script includes its Speaker 0: label. ASR/alignment uses 40 seconds of speech; separation uses 40 seconds of stereo music.
• Irodori RF/MF uses codec_decode_chunk_steps=256. WASM32 model metadata arenas fit within its 4 GiB address space; browser GPU buffer limits are separate.
• LFM3-Audio uses the published Q8_0 backbone and Q8_0 mmproj with its F16 codec, WASM64, default model controls and the WASM sinc input resampler. TTS uses seed 1234 and the American English female voice.
• PocketTTS Small 18M uses F32 weights, the matching Alba voice state, seed 1234, automatic attention (resolved flash), F32 KV and bounded four-frame offline decoding.
• VibeVoice uses WASM64 with a 16 GiB host-memory ceiling, separate from GPU memory.
• YuE2 uses Q4 main weights, an F16 VAE, cot=off, eight NAR steps, semantic_max_tokens=1500 and vae_tile_frames=256. Generation stops at the token cap.
• RE-USE uses 10-second offline chunks with one-second overlap. Realtime RE-USE uses streaming mode, exit layer 8 and zero lookahead.
• Supertonic's Q8-labelled package contains F32/I64 tensors; the table shows the actual types.

Output checks

PocketTTS Small 18M and LFM3-Audio TTS produced finite, nonconstant 24 kHz audio. Their warmed outputs were identical across repeats and matched the independently checked first-request outputs: Whisper Small transcribed all 43 input words correctly after English normalization (0% WER). LFM3-Audio ASR returned the same English transcript across all four requests.

RE-USE and Realtime RE-USE produced finite, nonconstant 24 kHz audio with the input sample count; waveform cosine similarity against matched native CUDA runs exceeded 0.9999997. Realtime RE-USE's streaming result validates incremental inference and final audio collection, not progressive browser playback or TTFA.

All generated TTS/music WAVs were finite and nonconstant. Independent SenseVoice transcription matched seven of twelve other TTS clips after normalization: Kokoro, VoxCPM, OmniVoice, Ming3, Qwen3-TTS, Irodori RF and VibeVoice. PocketTTS, Magpie and Irodori MF had small transcript differences; Supertonic and MOSS Nano had larger differences. These checks are not a model-quality benchmark.

Alignment produced 101 ordered words within the 40-second input. Separation produced both BS-RoFormer stems and all four HTDemucs stems.

See the WebGPU build guide to run models with your own weights and inputs.