TTS Voice-Generation Benchmark

This benchmark measures KokoroTTS end to end through the public /tts/generate HTTP endpoint. examples/voices.js is the single source of benchmark voices and language-specific input text, so the benchmark automatically follows the examples page as voices are added or corrected.

An official run performs these steps:

1. Load every voice and its localized introduction text from examples/voices.js.
2. Generate one unmeasured warmup for the first voice in each language.
3. Generate WAV audio five times for every voice, rotating through all voices once per round.
4. Write raw results and overall, per-language, and per-voice aggregates to results/latest.json.
5. Append the overall result to BENCHMARKS.md and detailed language/voice tables to DETAILS.md.

Start the application, then run:

task benchmark-smoke
task benchmark-tts BENCHMARK_COMMENT="baseline"

benchmark-smoke checks the runner with two voices and one measured generation each without changing benchmark history. benchmark-tts measures all 56 current voices five times each, following one warmup request per language.

Useful overrides:

task benchmark-tts BENCHMARK_REPETITIONS=2 BENCHMARK_LIMIT=9
task benchmark-tts GPU_DEVICE=1 APP_DEVICE=cuda:0 BENCHMARK_COMMENT="second host GPU"

Keep the same image, host GPU, device selection, voice source, and repetition count when comparing runs. These numbers measure generation performance; voice quality still requires listening tests.