LocalLLaMA asks for Mac speech stack

This is less a product launch than a sharp signal of where local voice AI still breaks down for developers: speech components are getting good, but the full realtime stack is not yet turnkey on-device.

• –The benchmark in the thread is already high: OpenAI Realtime and Google Live are praised for low latency and Indian language coverage, so local alternatives are being judged against production-grade cloud UX.
• –Sarvam is a credible anchor for this discussion because its speech stack is explicitly built around Indian languages, and its recent on-device work suggests local STT/TTS for this market is becoming practical.
• –The hardest missing piece is the middle of the pipeline: a compact, streaming-friendly local LLM that can respond fast enough on Apple Silicon without falling apart on multilingual prompts.
• –The post also highlights why developers keep chasing true speech-to-speech models: every extra handoff in a cascade adds latency, error propagation, and engineering complexity.
• –A 128GB Mac Studio is strong enough to make this a real deployment question rather than a toy experiment, which makes the thread useful as a demand signal for local voice infrastructure.