VoiceBox

A free, open-source text-to-speech application running entirely on local hardware. No cloud, no subscriptions, no email signup. Available on Mac, Windows, and Linux. Positioned as the open-source alternative to ElevenLabs.

Models

• Qwen-3 (0.7B parameters): Higher audio quality, best overall model for voice generation
• Chatterbox: Lower audio quality but supports slash-command effects: /laugh, /chuckle, /gasp, /cough, /sigh, /groan, /sniff, /shush, /clear_throat

Key Features

• Voice cloning: Record ~30 seconds of audio with reference text → creates a voice profile. Sensitive to audio quality — needs good mic and gain staging.
• Sound effects: Built-in presets (robotic, radio, echo chamber, deep voice). Effects are combinations of: low pass, high pass, chorus, reverb, delay, compressor, pitch shifter.
• Stories: Folder system for organizing and generating longer-form audio content.
• API: Local server with optional network access for automation integration.
• Multi-language: Depends on the TTS model selected (Chatterbox Turbo is English-only).

Hardware

Runs on Apple Silicon (M4 Max demonstrated) and other capable machines. Speed depends on local hardware. Models in the 7B range run well.

Rough Edges (as of March 2026)

• Custom effect presets can’t be saved (bug)
• Voice cloning fails on clipped audio
• No input device selector (uses system default)
• Chatterbox audio quality noticeably lower than Qwen-3

Security Note

Voice cloning quality is not yet good enough to fool family members, but improving rapidly. Dave Swift recommends setting up verbal passwords with loved ones as a precaution against voice spoofing attacks.

See Also

• Dave Swift — source
• llama.cpp — inference engine likely powering the local models
• Source: VoiceBox TTS