Open Source STT LLM

what free open source github repos are there that have speech to text you can feed to a large language model to summarize and notate live content like meetings

Below is a quick-scan of the most active **free, open-source GitHub projects** that already wire together speech-to-text and an LLM so you can capture, summarise and annotate live meetings. I grouped them by “how much they give you out-of-the-box” versus “how much you assemble yourself”. --- ### 1 Self-hosted desktop assistants (works on Mac/Win today) | Project | STT stack | LLM link-up | Why people like it | |---|---|---|---| | **Meetily** (`Zackriya-Solutions/meeting-minutes`, ★6.6 k, MIT) | Whisper.cpp compiled locally; selectable model sizes (tiny→large-v3-turbo) | Pluggable providers (Ollama, Groq, Anthropic). Generates real-time transcript, live diarisation and final action-item summary | Everything runs on your laptop, no external server; Homebrew, Docker or manual install. Linux build in progress. citeturn5view0turn7view0 | --- ### 2 Browser-extension “copilots” that join Google Meet / Teams for you | Project | Notes | |---|---| | **Amurex Meeting Copilot** (`thepersonalaicompany/amurex`, ★2.6 k, AGPL-3) – Chrome extension streams audio from the tab, does real-time suggestions, late-join recaps, full transcripts and follow-up-email drafts. Self-host to avoid the store. citeturn12view0 | --- ### 3 Cloud-native stacks (hosted on your own AWS/Azure etc.) | Project | Infra | Highlights | |---|---|---| | **Live Meeting Assistant (LMA)** (`aws-samples/amazon-transcribe-live-meeting-assistant`, MIT-0) | Everything deploys via AWS CDK/CloudFormation: Transcribe streaming → DynamoDB/AppSync → Bedrock LLM for auto-summary | Browser extension streams stereo audio; speaker names injected; produces searchable knowledge-base you can chat with afterwards. citeturn4view0 | | **LinTO AI** (multiple repos under `linto-ai/…`, AGPL) | Docker-Compose or Ansible deploy of micro-services: `LinTO-STT` (Whisper/Kaldi), `LinTO-Diarization`, `LLM-Gateway` for rolling summaries | Also gives you LinTO Studio for collaborative subtitle editing and a “bot” that joins video-conferences to feed the pipeline. citeturn9view0 | --- ### 4 Light-weight reference apps & starter templates | Repo | One-liner | |---|---| | **AI-Powered-Meeting-Summarizer** (`AlexisBalayre/AI-Powered-Meeting-Summarizer`, MIT) – Gradio front-end; whisper.cpp + Ollama; drop an audio file and get transcript + summary. Good minimal example to copy. citeturn8view0 | | **LLM-Minutes-of-Meeting** (`inboxpraveen/LLM-Minutes-of-Meeting`, MIT) – Flask app that turns audio/video into text, then creates editable MoM. Phase 2 roadmap adds real-time streaming. citeturn10view0 | --- ### 5 Building-block libraries for “roll your own” * **whisper-timestamped**, **whisper-streaming**, **WhisperLive**, etc. add low-latency streaming to Whisper models (useful if you need sub-second captions). citeturn15search2turn15search5 * Realtime WebSocket wrappers and React hooks (e.g., `whisper-live`) make it trivial to pipe mic audio straight into a Whisper model in the browser. citeturn15search1 --- ### Picking the right starting point * **Want “just works” on your laptop?** → Meetily. * **Need a Chrome-extension sidekick for Meet/Teams with no SaaS lock-in?** → Amurex. * **Already an AWS shop or need multi-user, searchable archives?** → Live Meeting Assistant. * **Media-studio workflows (journalists, subtitles) or multi-language rooms?** → LinTO AI. * **Prototyping or embedding into your own product?** → Grab one of the lightweight templates or the streaming libraries. All of the above are permissively licensed (MIT/AGPL/MIT-0) and regularly updated, so you can fork and adapt them to slot in your own local LLM or prompt-engineering tricks.

now write a paragraph on how you would do speech to text by coding but do not supply any code just one paragraph

To implement speech-to-text in your own codebase, you’d start by capturing raw audio—either live from the microphone or buffered from a file—then feed the PCM stream into an open-source inference engine like whisper.cpp or Vosk that runs a pre-trained acoustic-language model locally; you’d configure it for low-latency streaming so chunks of audio are transcribed as they arrive, optionally apply voice-activity detection and speaker-diarization to segment speakers in real time, and finally funnel the incremental text output into your downstream logic—whether that’s a websocket pushing live captions to the UI, a queue feeding a large-language-model prompt for summarization, or a datastore for searchable meeting archives—wrapping the whole pipeline with error handling, model-select toggles, and prompt templating so you can swap transcription or LLM back-ends without touching the surrounding application.