Explore Voicebox (Qwen3-TTS) for custom podcast voices #6
Description
Context
We previously tried ElevenLabs for custom voice synthesis for the podcast and hit issues. Found an open-source alternative worth exploring:
Voicebox - https://github.com/jamiepine/voicebox
- Open-source voice synthesis studio powered by Qwen3-TTS
- Could replace NotebookLM's default voices with custom ones
- Local/self-hosted = no API limits or recurring costs
Why This Matters
- NotebookLM gives us no control over voice selection or style
- ElevenLabs had quality/reliability issues when we tried it
- Open-source means we can iterate on voice quality locally
- Could enable consistent "host" voices across all episodes
To Investigate
- Verify repo exists and is actively maintained (link summary suggests it may not be live yet)
- Evaluate Qwen3-TTS voice quality vs ElevenLabs vs NotebookLM
- Test with a sample episode transcript
- Assess hardware requirements (can it run on our Mac?)
- Determine if it supports two-host conversational format
- Compare latency and generation time for 30-40 min episodes
- Evaluate voice cloning capabilities (custom host voices)
Integration Path
If viable, the workflow would be:
- Generate script/transcript (current workflow up to report.md)
- Use Voicebox to synthesize audio with custom voices
- Replace NotebookLM step entirely
This would give us full control over the audio generation pipeline.