Skip to content

Benchmarks

In-browser performance depends on your GPU, your CPU and your browser, so this page does not publish a leaderboard of numbers we collected on our machines. Instead: what the metrics mean, how they are measured, and a curated table of published measurements from external sources — each attributed. For numbers on your hardware, run the live demos on the model pages.

What is measured, and how

Real-time factor (TTS)

Generation time divided by the duration of the produced audio. RTF 1.0 means real time; 0.105 means speech is produced about 9.5x faster than it is played. Every TTS demo on this site displays live generation metrics as you run it.

Tokens per second (LLM)

Decode throughput of the voice agent's language model stage. The in-browser reference point is the WebLLM paper's measurement of Llama-3.1-8B at 41.1 tok/s on an M3 Max, about 71% of native Metal speed on the same machine.

Word error rate (STT)

The standard accuracy metric for speech-to-text: edits needed to turn a transcript into the reference text, lower is better. Cross-model WER comparisons are best taken from the Open ASR Leaderboard and each model's published card — not re-measured per visitor, because a single clip proves nothing about accuracy.

STT accuracy references we link to per model: the Open ASR Leaderboard and each model's Hugging Face model card, linked from every model page.

Published measurements (external)

External sources

These are external, published measurements — not run by TTSLab, and not reproducible on your machine without its exact context. Each row names its source. To see what a model does on your hardware, run its live demo.

Model / systemMetricPublished resultContextSource
Kokoro-82M — WebGPU backendReal-time factor (RTF)0.105 — about 9.5x faster than real timeApple M1 Max, in-browserbriantung.me, Aug 2026
Kokoro-82M — single-thread WASM (q8)Real-time factor (RTF)2.002 — about half of real timeApple M1 Max, in-browser (same machine as the WebGPU row)briantung.me, Aug 2026
Llama-3.1-8B via WebLLMDecode speed41.1 tokens/sec — roughly 71% of native Metal performanceApple M3 Max, in-browserWebLLM paper (arXiv 2412.15803)
Whisper tiny / base via whisper.cppTranscription speedRoughly 2-3x faster than real time on modern CPUsWASM / CPU, in-browserwhisper.cpp README
MoonshineCPU transcription latencyPaper claims 5-15x faster than comparably accurate Whisper models on CPUCPU, on-deviceMoonshine paper (arXiv 2410.15608)

Planned: a standardized suite

A reproducible in-browser benchmark suite — fixed test phrases for TTS, fixed clips for STT, results anchored to your device profile — is planned. Behind the scenes, generations on this site already pass through an automated QA harness that checks each render for truncation, looping, clipping and silence before it counts as a success. Subscribe below if you want to know when the public leaderboard lands.