TTS Compare

Overview

TTS Compare lets you generate the same input across multiple open-source text-to-speech models in a single run, then compare quality, latency, and per-model feature support side-by-side. The project addresses a recurring practical question, "which TTS should we ship?", by replacing ad-hoc one-off testing with a reproducible harness.

Models covered

The harness currently runs eight models spanning the modern open-source TTS landscape:

Architecture decisions

• Per-model isolated .venv. Each model gets its own environment under models/<name>/ to avoid CUDA/torch version skew between systems with conflicting requirements.
• JSON-over-stdin worker protocol. Each model is invoked as a subprocess that accepts a JSON request on stdin and writes a WAV file. Decoupling the harness from the model runtimes keeps the Textual TUI snappy and lets a model crash without taking the rest of the run down.
• Hardware adaptive. Detects CUDA, MPS, or CPU and configures models accordingly.
• Three-screen TUI. Input → model selection → execution, with real-time logs streaming as each model generates speech.

References

• Repository: github.com/andrewmarconi/tts-compare