LM Studio 1.1.3 brings local voice transcription to Linux

LM Studio 1.1.3 adds real-time, on-device speech transcription to Linux, joining the existing macOS and Windows implementations. The desktop app can now convert microphone audio into prompt text without sending recordings to a cloud transcription service.

LM Studio provides a graphical interface and an OpenAI-compatible local server for running large language models on personal hardware. Voice input previously required a separate speech-to-text service or community software such as this community bridge. Native integration removes that extra process and its supporting code.

What shipped in 1.1.3

The 1.1.3 release notes describe local voice transcription as the main Linux addition. The update also includes several changes to model execution and chat behavior:

• Local transcription: Real-time microphone input is available on macOS, Windows, and Linux.
• Hardware acceleration: Apple Silicon and NVIDIA GPUs are supported, while AMD GPU support remains in development.
• Inline images: Images created by agents can render directly inside conversations.
• MTP speculative decoding: More compatible models can use a technique that verifies multiple candidate tokens to accelerate generation.
• Runtime updates: The release adds llama.cpp 2.38.0 extension packs and smoother handling of remote models through LM Link.

LM Studio has not identified the speech model or inference engine behind the feature. Its release materials also omit supported languages, model size, memory use, latency measurements, accuracy results, and audio retention details.

Hardware support at a glance

Platform	Hardware	Published status
macOS	Apple Silicon	Available with acceleration
Windows	NVIDIA GPU	Available with acceleration
Linux	NVIDIA GPU	Added in version 1.1.3
Windows or Linux	AMD GPU	In development
Other configurations	Intel Mac, integrated GPU, or CPU-only	No accelerated support announced

Why latency depends on the whole machine

Real-time transcription processes short audio segments while the speaker is still talking, producing partial text and revising it as more audio arrives. That workload must share compute and memory with the language model generating the response.

Apple Silicon uses a unified memory pool for both workloads. NVIDIA systems divide the available VRAM among the speech model, language model, context cache, and other active processes. A large language model can therefore slow transcription or trigger memory pressure even when the speech component runs efficiently.

No official benchmark establishes expected latency for this release. Performance will depend on the GPU, available memory, selected language model, prompt length, audio quality, and any agent tools running during the conversation.

Where local speech input fits

• Coding and note-taking: Developers can dictate prompts, documentation, or rough notes directly into a local model.
• Sensitive audio: Local processing reduces exposure to third-party transcription providers, although regulated deployments still require security and compliance review.
• Offline systems: Voice input can work on disconnected, air-gapped, or unreliable networks.
• Multilingual workflows: Local transcription can remove per-minute API charges when the bundled engine supports the required language.
• Agent interfaces: Spoken instructions can enter the same chat flow used for tools, images, and local documents.

How to try it

Bionic 1.1.3 is a free update for existing users. The built-in updater can install it, while new users can obtain the application from LM Studio downloads.

1. Install version 1.1.3 or later.
2. Grant LM Studio access to the system microphone.
3. Load a compatible local language model.
4. Select voice transcription in the chat composer.
5. Speak, review the generated text, and submit it as a prompt.

LM Studio’s existing OpenAI-compatible server remains available at http://localhost:1234/v1, so applications already sending text prompts to that endpoint can continue doing so. The voice feature converts speech into text inside the desktop chat interface.

The release notes do not document an OpenAI-compatible audio endpoint such as /audio/transcriptions. Developers building automated speech pipelines should therefore treat transcription as an application feature until LM Studio publishes an audio API.

Privacy stops at the transcription boundary

On-device speech recognition removes the cloud transcription hop, but the generated text can still leave the computer through remote models, LM Link, agent tools, plugins, networked MCP servers, backups, or application telemetry. Deployments handling confidential material should audit the full request path, chat storage, logs, microphone permissions, and every enabled integration.

Accuracy also depends on the undisclosed speech model and the recording environment. Background noise, overlapping speakers, specialized vocabulary, and unfamiliar accents can increase errors. The release covers speech-to-text; a fully spoken assistant still requires a separate text-to-speech component.

Voice joins the local model stack

By folding microphone input into the same desktop runtime as chat, vision, tools, and local model serving, LM Studio reduces the number of services required for a private voice interface. Version 1.1.3 gives Linux users the integrated workflow now, while AMD acceleration, documented audio APIs, and native speech output remain open areas for future releases.