Local-first AI vs cloud AI — which is winning for voice generation?

Most voice AI services — ElevenLabs, PlayHT, Murf — run in the cloud. You upload your text, they generate audio, you download it. Per-character pricing.

But there's a clear shift toward local-first AI happening across the board. Apple's MLX framework, Ollama for LLMs, Whisper.cpp for transcription. Models are getting small enough and hardware is getting fast enough that "run it on your own machine" is a real option.

For voice generation specifically, the tradeoffs are interesting:

Cloud advantages:

• No hardware requirements

• Always the latest model

• Instant setup, no installation

Local advantages:

• Privacy — scripts never leave your machine

• No per-character costs (generate as much as you want)

• Works offline

• No usage caps

• Iterations are free (change a word, regenerate just that chunk)

The performance gap is closing fast. On Apple Silicon, local TTS can generate at 6x real-time — a 10-second clip in under 2 seconds. Quality that a year ago required cloud-only models is now running on laptops.

Where do you think this goes?

• Will cloud TTS pricing hold, or will local models commoditize voice generation the way Whisper commoditized transcription?

• For enterprise use cases (e-learning, training, compliance), does the privacy argument for local processing outweigh cloud convenience?

• Is there a hybrid model that makes sense — cloud for occasional use, local for heavy production?

I've been building in the local-first space for the past year and I'm biased, but I think the shift is accelerating faster than most people expect. Curious what others are seeing.