Vertere

Fully local speech ↔ text web UI. No cloud calls. English only.

System prerequisites

• ffmpeg — audio extraction from video files (sudo apt-get install ffmpeg)
• NVIDIA GPU + driver (recommended) — falls back to CPU automatically
• Python 3.11+ with uv

Install & run

uv sync
uv run python app.py

Open http://localhost:7860.

CUDA notes

torch 2.11 ships CUDA 13 runtime libs, but ctranslate2 (used by faster-whisper) is built against CUDA 12. The nvidia-*-cu12 wheels listed in pyproject.toml provide the needed cuBLAS 12 / cuDNN 9 / CUDA runtime 12 libs, and stt.py preloads them via ctypes before torch is imported. If CUDA fails at runtime, the pipeline falls back to CPU automatically.

Usage

Input	Action
Audio file (.wav, .mp3, .m4a, .flac, .ogg)	Transcribe to text
Video file (.mp4, .mov, .mkv, .webm)	Extract audio → transcribe
Text/Markdown file (.txt, .md)	Synthesize to speech
PDF (.pdf)	Extract text → synthesize to speech
Pasted text	Synthesize to speech

TTS engines

Engine	Notes
Kokoro-82M	Fast, lightweight. Multiple voices. No reference audio needed.
F5-TTS	Higher quality. Uses Kokoro-generated reference audio on first run.

First run

Models download automatically on first use (~3 GB for Whisper large-v3, ~300 MB for Kokoro-82M, ~1.2 GB for F5-TTS v1 Base).