Speech models by hardware
For each hardware class, these are the models we've measured or that have a sourced, device-specific statement. We don't make pages for pairs without evidence. Use the picker for everything else (it also shows inferred fits, clearly labelled).
Raspberry Pi 5
Hobbyist single-board computer. Common for Home Assistant voice and offline assistants. Small models only. Typical examples: Raspberry Pi 5 8 GB. Memory assumed: 8 GB RAM.
- STT Whisper tiny Sourced statement
- STT Whisper base Sourced statement
- STT Moonshine Streaming Medium Sourced statement
- TTS Piper (voices) Sourced statement
- TTS Pocket TTS Sourced statement
- TTS Supertonic 3 Sourced statement
- TTS NeuTTS Air Sourced statement
CPU only (Intel N100 mini-PC / x86 desktop)
No GPU. Covers low-power N100 boxes up to desktop CPUs. Our own CPU numbers come from an 8 vCPU cloud Xeon, which is faster than an N100. Typical examples: Intel N100 mini-PC 16 GB, any x86 laptop/desktop without a usable GPU, cloud CPU VM. Memory assumed: 16 GB RAM.
- STT Whisper tiny Measured by us
- STT Whisper base Measured by us
- STT Whisper small Measured by us
- TTS Kokoro 82M Measured by us
- TTS Piper (voices) Measured by us
- TTS Pocket TTS Sourced statement
- TTS Supertonic 3 Sourced statement
- TTS NeuTTS Air Sourced statement
- TTS LuxTTS Sourced statement
Mac (Apple Silicon M1–M4)
Unified memory shared by CPU and GPU. Runtime support varies a lot (Metal via whisper.cpp/MLX; PyTorch MPS is often partial). Typical examples: MacBook Air M2 16 GB, Mac mini M4 16 GB. Memory assumed: 16 GB unified memory.
- STT Whisper large-v3-turbo Sourced statement
- STT Whisper small Sourced statement
- TTS Kokoro 82M Sourced statement
- TTS F5-TTS Sourced statement
- TTS Pocket TTS Sourced statement
- TTS NeuTTS Air Sourced statement
NVIDIA GPU, 8 GB VRAM
Entry gaming GPU. Fine for most small/medium speech models; tight for 3B+ models in fp16. Typical examples: RTX 3060 Ti, RTX 3070 / 3070 Ti, RTX 4060. Memory assumed: 8 GB VRAM.
- STT faster-whisper (large-v3, CTranslate2) Sourced statement
- TTS LuxTTS Sourced statement
NVIDIA GPU, 12–16 GB VRAM
Mid-range. Most open TTS/STT models fit in fp16; LLM-based TTS at 3–4B may need 16 GB. Typical examples: RTX 3060 12 GB, RTX 4070 / 4070 Ti, RTX 4060 Ti 16 GB, Tesla T4 16 GB. Memory assumed: 12–16 GB VRAM.
- STT Voxtral Mini 4B Realtime (2602) Sourced statement
- TTS Pocket TTS Sourced statement
NVIDIA GPU, 24 GB+ VRAM
Enthusiast/prosumer. Runs nearly every open speech model, including 8B-class TTS. Typical examples: RTX 3090, RTX 4090, RTX 5090 (32 GB). Memory assumed: 24 GB VRAM.
- STT Voxtral Mini 4B Realtime (2602) Sourced statement
- TTS NeuTTS Air Sourced statement
- TTS IndexTTS-2 Sourced statement
- TTS Miso TTS 8B Sourced statement