Guides · Updated

Local speech-to-text and text-to-speech on your PC

Transcribe recordings with Whisper and generate natural-sounding speech with Piper or Kokoro, all offline on your own computer, plus how to add voice to your local AI chat.

Not sure which model to pull? Use the GPU model picker: enter your VRAM and RAM and get a list of Ollama models that fit.

Voice is one of the easiest things to move off the cloud. Open speech models are small compared with chat models, many run fine on a CPU, and your recordings never leave your machine. Ollama is built around text chat, so speech uses separate tools that sit alongside it.

Speech-to-text: Whisper

Whisper is OpenAI’s open-source speech recognition model, and most local transcription tools are built on it. It handles many languages, adds punctuation, and can translate speech into English.

Pick a model size

ModelParametersVRAM (official Python version)Notes
tiny39M~1 GBFastest, least accurate
base74M~1 GBFine for clear speech
small244M~2 GBA good middle ground
medium769M~5 GBNoticeably more accurate
large1550M~10 GBMost accurate, slowest
turbo809M~6 GBClose to large, much faster; transcription only, no translation

tiny to medium also come in English-only versions (base.en, small.en and so on), which tend to be more accurate for English, especially at the smaller sizes. Other implementations like whisper.cpp and faster-whisper need less memory than the figures above. Why memory matters: VRAM and quantisation explained.

The easy way: a desktop app

  • Buzz: an open-source app for Windows, Mac and Linux. Drop in an audio or video file and get a transcript or subtitles.
  • Subtitle Edit (Windows): a subtitle editor that can generate subtitles with Whisper.

The command-line way

Install Python and FFmpeg (on Windows: winget install Gyan.FFmpeg), then:

pip install -U openai-whisper
whisper meeting.mp3 --model turbo --output_format txt

The model downloads on first use, then everything runs offline. Use --language to skip language detection, and --output_format srt for subtitles. It runs on the CPU if you don’t have an NVIDIA GPU, just more slowly.

Faster engines

  • whisper.cpp: a C/C++ port that runs well on CPUs and supports NVIDIA, AMD and Intel GPUs via CUDA or Vulkan, and Apple Silicon via Metal. Prebuilt Windows builds are on its GitHub releases page.
  • faster-whisper: a Python reimplementation that its authors report is up to 4x faster than the official version, with lower memory use. Many apps, including Open WebUI, use it under the hood.

Text-to-speech: Piper, Kokoro and friends

ToolStrengthHardware
Windows natural voicesBuilt in, reads your screen aloud via NarratorWindows 11 (22H2 or later)
PiperVery fast, many languages and voicesRuns well on a CPU, even a Raspberry Pi
KokoroVery natural for its size (82M parameters)CPU is fine, GPU is faster
Voice-cloning models (e.g. Chatterbox, XTTS-v2)Speak in a voice from a short sampleA GPU helps a lot
  • Windows natural voices: Windows 11 lets you download more natural voices for Narrator (Settings → Accessibility → Narrator). Once downloaded they work offline, but only inside Narrator, not as a general voice for other apps.
  • Piper: a lightweight neural voice engine, widely used for local voice assistants in Home Assistant. Install with pip install piper-tts and download a voice. The command syntax has changed between versions, so follow the project’s current README.
  • Kokoro: an open-weight model with Apache-licensed weights that sounds natural for its size. Community projects wrap it in a local server with an OpenAI-compatible API, which makes it easy to plug into chat apps.
  • Voice cloning: only clone voices you have clear permission to use, and check each model’s licence before using it commercially. Some popular ones, such as XTTS-v2, have non-commercial licences.

Add voice to your local AI chat

Open WebUI can do both directions. In Admin Panel → Settings → Audio:

  • Speech-to-text: choose the local Whisper option. It runs inside Open WebUI, and the model downloads the first time you use it.
  • Text-to-speech: the default uses your browser’s voices. Some browsers offer online voices alongside local ones, so for fully offline speech point it at a local TTS server, such as a Kokoro server with an OpenAI-compatible API.

Then use the microphone button to dictate, or Standard call mode for a back-and-forth voice conversation with your local model. Leave Call mode on Standard: Realtime sends the audio to OpenAI or Azure. To connect those pieces into a spoken assistant, and to keep the microphone off cloud speech services, see Build a private voice assistant on your own PC.

Tips

  • Clean audio beats a bigger model. A decent microphone and a quiet room help more than going from small to medium.
  • Long recordings: let them run in the background, or use whisper.cpp or faster-whisper for speed.
  • Sharing a GPU with a chat model: Whisper and your chat model both need VRAM. On an 8 GB card, run Whisper on the CPU or use a smaller chat model. The GPU picker shows what fits.
  • On a laptop: plug in for long transcriptions. More in Running local AI on a laptop.
  • Privacy: transcripts of meetings are often more sensitive than the documents. Store them as carefully as anything else; see Local AI for privacy, step by step. Which voice settings send audio out, and how to turn them off: What leaves your PC?.

Sources: Whisper on GitHub, whisper.cpp, Open WebUI docs.

Get the Local AI weekly

New models worth running, what fits which GPU, and setup fixes. One short email a week, free. Unsubscribe any time.

Subscribe free →

Lila Forge Kits

Local-LLM App Starter

A tested starter for building your own app on Ollama: chat, RAG and a GPU-fit checker. Skip the setup hours.

kits.lilaforge.com →
VibeEngine

Best local LLM tools, ranked

Ollama, LM Studio, Open WebUI and llama.cpp scored from public evidence, with top picks.

vibeengine.lilaforge.com →