Local voice and dictation: goodbye to the monthly fee

Dictation and speech synthesis were until recently subscription territory (Wispr Flow, Superwhisper). In 2026 open source reached a very usable local level: Whisper remains the transcription standard, VoxCPM adds multilingual synthetic voice with cloning, and FluidVoice offers on-device dictation on Mac.

For realtime voice agents, the focus moved to frameworks: Pipecat and LiveKit orchestrate transport, recognition and synthesis to build assistants that listen and respond. That is the base serious conversational agents are built on.

The honest nuance from the reviews: for pure naturalness, services like ElevenLabs still win, but the gap is narrowing and, for most internal cases, local already does the job without sending audio to third parties.