Voxcast

Give it a voice. Give it words. Get speech.
Voxcast clones a voice from a short sample, then turns any text into speech in that voice—all on your machine, no GPU required. Add Podcast mode and you get AI personas debating each other in cloned voices, streamed as they’re generated.

---

What it is

• Voice clone TTS — Upload a reference clip (WAV or MP3, a few seconds of clear speech). The app extracts a voice embedding, then you type text and generate speech in that voice.
• Podcast mode — Create personas (name, role, prompt), optionally attach a voice to each, and run a streaming debate: the AI speaks, you hear it in the right voice.
• Local-first — FastAPI backend + LuxTTS/ZipVoice on CPU. First run downloads the model from Hugging Face; after that it’s all offline.

---

Quick start

You need: Node.js, uv. For Podcast mode, a Backboard API key in .env.

./start.sh

That clears frontend cache (not .env or data), syncs backend deps, starts the API on http://localhost:8001, and runs the UI with /api proxied. Open http://localhost:5173 in the browser. Use Voxcast for voice clone, Podcast for debates.

---

How it works

Voice clone

1. Upload reference → backend stores it and runs a job to compute a voice embedding (LuxTTS on CPU).
2. Frontend polls /api/jobs/{job_id} until the job finishes → you get an embedding_id.
3. You type text and hit Generate → frontend sends text + embedding_id to POST /api/tts → another background job runs TTS.
4. UI polls the TTS job → then fetches /api/audio/{audio_id} and plays (or downloads).

So: reference → embedding job → embedding_id; then text + embedding_id → TTS job → audio.

Podcast mode

• Personas live in Backboard; each can have an optional voice sample. The backend uses that to synthesize that persona’s lines.
• You start a session with a topic and personas. The backend streams SSEs: “turn” (who + text) or “audio” (segment id). The frontend fetches each segment from /api/podcasts/{id} and plays in order so the debate runs with the right voices.

Backend in brief

• FastAPI in backend/: voice (upload reference, create embedding), tts, jobs (poll), files (serve audio/podcasts), pod (personas, history, stream), model, health.
• LuxTTSService loads LuxTTS from Hugging Face at startup; creates embeddings and generates WAV (chunking + ONNX, CPU-only).
• Backboard for podcast personas/sessions; optional if BACKBOARD_API_KEY is unset.
• Temp files under a configurable dir; cleaned on shutdown.

---

Project layout

Secrets and config go in .env—don’t commit it. Backend: uv + pyproject.toml; frontend: npm.