MTS

Meeting Transcriber & Summarizer (FFmpeg v8 Whisper + LM Studio / Ollama )

🇫🇷 Lire ce README en français

A tiny Gradio app that records from your browser microphone or uploads an audio file, transcribes it locally using FFmpeg v8 + the whisper filter (with Whisper.cpp models), and sends the transcript to a local LLM (LM Studio) to produce a clean, structured meeting minutes (Markdown).

Highlights

• One-click record or upload (Gradio 4.x Audio).
• Robust capture: pre‑roll (to avoid missing the first words), optional VAD, and input normalization to WAV 16 kHz mono.
• Flexible output: text / srt / json.
• Built‑in prompt template CRUD persisted to prompt_templates.json and used as system prompt for LM Studio.
• Persistent UI options (language 🇬🇧/🇫🇷 and light/dark theme) saved to ui_settings.json.
• New Transcriptions tab with full CRUD to revisit transcripts and their summaries.

---

Architecture (at a glance)

1. Gradio UI (microphone/upload → Audio component).

2. FFmpeg 8 + whisper filter (with Whisper.cpp ggml model) → transcript file.

   • Optional pre‑roll (adds a short silence at the start) and VAD.
   • Audio is normalized to WAV 16 kHz mono for reliability.

3. LM Studio (OpenAI‑compatible API) → structured Markdown meeting minutes. You can use Ollama too.

---

Requirements

• Python 3.10+

• FFmpeg 8.0+ compiled with the whisper filter (requires Whisper.cpp). Verify with:

  • Linux/macOS: ffmpeg -hide_banner -filters | grep whisper
  • Windows: ffmpeg -hide_banner -filters | findstr whisper

• LM Studio running in local server/developer mode (OpenAI‑compatible), or Ollama, usually at http://localhost:1234.

• Whisper.cpp ggml model file(s), e.g. ggml-large-v3-turbo.bin.

Python packages:

gradio>=4.44.1
uvicorn>=0.30
starlette>=0.37
anyio>=4.4
h11>=0.14
httpx>=0.27
httpcore>=1.0
python-dotenv

Privacy: All audio processing happens locally via FFmpeg; the transcript is summarized by your local LM Studio instance.

---

Installation

1. Clone the repo and enter it.

   git clone https://github.com/magicmars35/AutoTranscriptReport
   cd AutoTranscriptReport

2. (Optional but recommended) Create a venv and activate it.

   For Windows :

   python -m venv venv
   venv\Scripts\activate

3. Install deps:

   pip install -r requirements.txt

4. Place a Whisper.cpp model under ./models/, e.g. ./models/ggml-large-v3-turbo.bin. Models can be found here : https://huggingface.co/ggerganov/whisper.cpp/tree/main

5. Ensure FFmpeg 8 supports the whisper filter (see Requirements above).

6. Start LM Studio (or Ollama) in server mode (Developer tab → Start server). Default base URL is http://localhost:1234 and the API path is usually /v1/chat/completions.

---

Configuration (.env)

Create a .env at the project root if you want to override defaults:

# LM Studio
LMSTUDIO_BASE_URL=http://localhost:1234
LMSTUDIO_API_PATH=/v1/chat/completions
LMSTUDIO_MODEL=Qwen2.5-7B-Instruct
LMSTUDIO_API_KEY=lm-studio

# FFmpeg + Whisper
FFMPEG_BIN=ffmpeg
WHISPER_MODEL_PATH=./models/ggml-large-v3-turbo.bin
WHISPER_LANGUAGE=fr

# Templates storage
TEMPLATES_PATH=prompt_templates.json

Defaults are the same as the example above. You can also edit the constants at the top of the Python file.

---

Run

python app.py

Gradio will print a local URL (e.g. http://127.0.0.1:7860).

---

Using the App

1. Record or Upload: Use the single Audio widget to record from the browser mic or upload an existing file (.wav, .mp3, .m4a, etc.).

2. Choose output format: srt (default) is great to visually verify timestamps; text for plain text; json for programmatic use.

3. (Optional) Advanced settings:

   • Pre‑roll (ms): add a short silence at the start so the very first words aren’t cut (try 250–500 ms).
   • Queue (ms): buffering window for VAD; larger values may help stabilize segmentation.
   • VAD (Silero): enable only if you provide a Silero VAD ggml model path; otherwise keep it off to preserve natural pauses.

4. Click Transcribe → FFmpeg runs the whisper filter and writes the transcript file into ./transcripts/.
5. Inspect the transcript/SRT/JSON shown in the UI.
6. Pick or edit a Prompt Template and click Summarize → LM Studio returns a structured Markdown meeting minutes.
7. (Optional) Use the Options tab to switch UI language (English/French) or theme (light/dark). Choices persist to ui_settings.json.

---

Prompt Templates (CRUD)

• Templates are persisted to prompt_templates.json in the project root.
• Each entry is a mapping of name ➝ prompt content.
• The UI provides buttons to Reload, Save, and Delete templates. The selected template’s content is sent as the system prompt.
• You can also hand‑edit prompt_templates.json while the app is stopped.

Example prompt_templates.json:

{
  "default": "You are an assistant specializing in meeting minutes...",
  "brief": "Write a concise summary focusing on decisions and actions."
}

The app trims the template content to a safe length before sending it to LM Studio.

---

UI Settings

• The Options tab lets you choose interface language (English or French) and theme (light or dark).
• Labels for each language are stored in strings/<lang>.json for easy translation.
• Selections are persisted to ui_settings.json so your preferences are restored on next launch.
• Delete or edit this file to reset the UI settings.

---

Model Choices

Place your chosen Whisper.cpp ggml model file under ./models and point WHISPER_MODEL_PATH to it. Recommendations:

• ggml-large-v3-turbo.bin (~1.5 GB): great quality/speed balance, multilingual.
• ggml-large-v3.bin (~2.9 GB): highest quality, heavier.
• ggml-medium.bin (~1.5 GB): good quality, lighter than large.
• ggml-small.bin (~466 MB): fast and decent for French.

All models are multilingual unless the filename ends with .en. Models can be found here : https://huggingface.co/ggerganov/whisper.cpp/tree/main

---

Tips & Troubleshooting

• Check whisper filter availability: if ffmpeg -filters does not list whisper, your build isn’t compatible. Install/compile FFmpeg 8 with Whisper support.
• Windows paths: the FFmpeg filter parser doesn’t like drive letters (D:) and backslashes. This app writes transcripts using relative POSIX‑style paths (slashes /) to avoid that.
• Microphone quirks / first words missing: increase Pre‑roll (e.g., 300–800 ms). Keep VAD off unless you really need it. The app also auto‑converts inputs to WAV 16 kHz mono for consistency.
• Only partial audio transcribed: try disabling VAD, increase queue (ms), and ensure your input isn’t corrupted.
• LM Studio errors (404/connection refused): make sure LM Studio’s local server is running and that LMSTUDIO_BASE_URL and LMSTUDIO_API_PATH match your version.
• Slow on CPU: prefer large‑v3‑turbo or small; quantized variants can help.

---

Roadmap (ideas)

• Export DOCX/PDF for the final minutes.
• Diarization / speaker labels.
• Multi‑language post‑processing and translation.
• Batch mode and watch folders.

---

License

MIT