Piper

Whisper covers listen. Without a local TTS pack, “voice round-trip” only exists as a stage name inside pipeline frameworks. Piper is the research default when spoken replies must stay on-device (privacy, offline Studio, Home Assistant-class boxes) and still bind to the same gate rules as text agents: speech is channel chrome; approve / spend / mutate stay elsewhere.

Text enters Piper; espeak-ng phonemizes; a neural voice model synthesizes audio. Operators pick a voice ONNX (or trained voice), call the CLI / Python API / HTTP server, and stream or write WAV/raw audio to a transport that Pipecat or LiveKit Agents already owns.

1. Install piper-tts (or build libpiper).
2. Download or train a voice model.
3. Synthesize text → audio (CLI, Python, or HTTP).
4. Hand audio to the voice-agent transport; keep tool side effects on the gated path.

Related: whisper · pipecat · livekit-agents · vocode · topics/03-channels

• Use when: you need local neural TTS as the speak half of a voice agent (offline, privacy, or HA-style boxes) and already have STT + agent loop elsewhere.
• Skip when: you only need a hosted voice API with no local weights, or you need a full telephony agent product — prefer Vocode / LiveKit Agents as the product and keep Piper as optional synthesizer.
• Prefer instead: Whisper for STT; pipeline packs for STT→LLM→TTS composition; commercial TTS SaaS when local GPL voices are unacceptable.

• License: piper1-gpl is GPL-3.0 — embedding in proprietary closed agents has copyleft implications; confirm counsel before shipping.
• Maintainer signal: OHF is actively seeking maintainers; treat release cadence as an ops risk.
• Quality / language: voice coverage and naturalness vary by ONNX; measure barge-in and latency on your transport, not demo WAVs alone.
• Scope: TTS is not the ledger, not the approval authority, and not STT.