Audio#Speech to text

AudioNotes: Turn Audio and Video into Structured Notes, Locally

AudioNotes runs speech recognition locally with FunASR, turning transcripts into Markdown via a local Ollama model, keeping data on-device for meetings.

Project facts

GitHub Ecosystem
Repositorygithub.com/harry0703/AudioNotes
License
MIT
Language
Python
Stars
2,525
Data checked
2026-10-06

Snapshot figures reflect the check date and may change over time.

Turning a meeting recording into text usually means listening to it over and over. AudioNotes moves that whole process onto your own machine: upload an audio or video file, or record straight from the browser, and FunASR transcribes it locally first, then a local LLM running on Ollama organizes the transcript into a Markdown note. GitHubDaily recommended it in February 2025.

Core features

  • Audio/video to transcript: upload a file to transcribe, or record directly in the browser; the first run needs an internet connection to download a roughly 2.15GB speech model.
  • Notes and Q&A: the transcript is organized into structured Markdown by an LLM, and you can keep asking follow-up questions — who’s responsible, when, what was decided — with conversation history preserved.
  • Hot words and speaker separation: domain-specific hot words improve accuracy; enabling speaker separation in .env produces SRT subtitles tagged by speaker.
  • Local-first: recognition and Q&A run entirely on your machine, with no third-party AI calls by default, and the web service is only accessible locally by default.
  • Default model pairing: speech recognition uses FunAudioLLM/Fun-ASR-Nano-2512, and note organizing/Q&A uses qwen3.5:2b; to recognize more languages, change ASR_MODEL in .env to FunAudioLLM/Fun-ASR-MLT-Nano-2512 and restart — it doesn’t switch automatically.

Typical use cases

  • Meeting summaries: turn a recording into notes, then directly ask who owns what, by when, and what was agreed.
  • Interview and podcast archiving: journalists and researchers convert raw material into searchable Markdown archives.
  • Lecture review: convert lecture videos or voice memos into transcripts for quick review later.
  • Moving to a new machine: under Docker, the whole audionotes folder is all your data — stop the service, copy the folder over, install Docker and Ollama, pull the default model, and you’re back up.

Quick start

First install and start Docker Desktop and Ollama, then run the following in an empty directory:

ollama pull qwen3.5:2b
curl -fsSL https://github.com/harry0703/AudioNotes/raw/main/docker-compose.yml -o docker-compose.yml
curl -fsSL https://github.com/harry0703/AudioNotes/raw/main/.env.example -o .env
docker compose up -d --remove-orphans

Once it’s up, visit http://localhost:15433/ — the default username and password are both admin. The default image runs fine on CPU alone; on Linux with an NVIDIA GPU, you can enable GPU acceleration with docker-compose.gpu.yml as described in the repo. If you’d rather skip Docker, you can also run from source with Python 3.12, ffmpeg, and Ollama installed, starting the service with python -m chainlit run main.py.

Summary

This suits anyone who regularly organizes meeting, interview, or lecture recordings but doesn’t want to hand that content to a cloud AI service — for the opposite direction, turning text into audio locally, see ebook2audiobook. The default model combination is friendly to Chinese speech recognition, and Q&A runs entirely on local Ollama, so there’s no third-party AI API key to obtain. The project is MIT-licensed and still maintained, with its most recent update in August 2026. A few caveats: first use requires internet access to download images and models; Docker is recommended with at least 12GB of RAM (16GB for long recordings), plus around 10GB of free disk space; on a shared machine, change the default credentials in .env; and recognition quality is affected by background noise, accents, and overlapping speakers. As of October 6, 2026 the repo has 2,525 stars, with its most recent push in August.