Use Liquid AI’s LFM3-Audio-2B with llama.cpp to turn text into speech and transcribe recordings on Windows, Mac or Linux, with nothing leaving your computer.
Estimates for clips under a minute. Not measured by YouRunAI.
| Machine | LFM3-Audio-2B · Q4_K_M GGUF | LFM3-Audio-2B · Q8_0 GGUF |
|---|---|---|
| NVIDIA · 4 GB VRAM | Fits on the GPU | Fits on the GPU |
| AMD Radeon · 4 GB VRAM | Uses the Vulkan or HIP llama.cpp build, else CPU | Same runtime requirement |
| Intel Arc · 4 GB VRAM | Uses the Vulkan llama.cpp build, else CPU | Same runtime requirement |
| Apple Silicon · 8 GB unified memory | Comfortable | Comfortable |
| CPU only · 8 GB RAM | Works; speech output is the slowest step | Works; a little slower |
Download one small model, have it read a sentence aloud into a WAV file, then transcribe that file back to text. One model does both directions on your own computer.
Select the weights you want. The instructions, downloads and settings change with your choice.
The publisher has not provided a complete memory target for this task. Check its hardware guidance before starting.
Short clips run comfortably on a modern CPU. Liquid AI notes the speech decoder is the slowest part on x86 CPUs; Apple Silicon is fastest.
Choose where you want to run it. The instructions below update to match your selection.
LFM3-Audio support arrived in llama.cpp v0.8.0 (build b11960). In PowerShell, install or upgrade llama.cpp with winget, close and reopen PowerShell, and check the version.
llama.cpp v0.8.0 release (LFM3 support) Official llama.cpp releasesIn PowerShell, make a working folder and download Liquid AI’s Q4_K_M GGUF and the audio projector it needs (1.1 GB + audio projector). Both files are required for speech in and out.
Liquid AI GGUF filesRun llama-tts from the same folder. It reads the sentence aloud in the built-in male voice and saves speech.wav. Use --tts-voice female for the other voice. Play speech.wav to check it says the sentence clearly.
| Setting | Value |
|---|---|
| Voice | male (female is the default) |
| Language | en (default) |
Now have the model transcribe that file with Liquid AI’s published ASR command: press Enter to run it and the model transcribes speech.wav, printing the text in PowerShell. It uses the documented "Perform ASR in English." system prompt and greedy decoding (temperature 0). If the prompt stays open for another message afterwards, press Ctrl+C. Save the printed transcript as transcript.txt in the lfm3-audio folder so you can compare it. To transcribe your own recording, save a short WAV or MP3 in the folder and replace speech.wav with its name; use "Perform ASR." to let the model detect the language.
Publisher ASR commandRead transcript.txt and compare the transcript with the sentence you typed: the date and time should match exactly. Disconnect from the internet and repeat both commands to confirm everything runs on this computer. For a browser chat with audio attachments, llama-server with the same -m and -mm files serves a web page at http://127.0.0.1:8080.
speech.wav says the sentence you typed, and transcribing it prints the same sentence, including the day and time, with the internet disconnected.
Live, streaming voice chat uses llama.cpp’s experimental /v1/realtime API and is not covered here. Commands are Liquid AI’s published llama.cpp examples, checked November 9, 2026; not run by YouRunAI.
Your llama.cpp is older than v0.8.0. Upgrade it (winget upgrade llama.cpp or brew upgrade llama.cpp) and check the version again.
Keep the exact system prompt "Perform ASR in English." and the "<__media__>" prompt, and use --temp 0.
Convert it to WAV or MP3 first; phone apps often save M4A.
Original instructions, model files, and compatibility notes behind this setup.
Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.
Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.
Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.