A real 81-minute university lecture (archives law, in Italian, from archive.org, CC BY-NC-SA), straight from the link, on the free tiers. It took 5½ minutes from link to notes:

$ distillo "https://archive.org/download/Lezione6LegislazioneArchiviStefanoTwardzik/lezione_6_lezione_twardzik_10_11_09.mp3" -n 8 --lang it

# lezione_6_lezione_twardzik_10_11_09

**8 nozioni chiave** (su 53) · 1:21:02 · `gemini:gemini-3.6-flash, groq:openai/gpt-oss-120b`

2. **Definizione di versamento archivistico** `02:39` · _definizione_
   Il versamento rappresenta il trasferimento della responsabilità delle serie archivistiche dagli uffici dell'amministrazione attiva (statali, amministrativi e giudiziari) all'amministrazione archivistica che afferisce al Ministero dei beni culturali.

   > “I versamenti sono questi trasferimenti di responsabilità di serie archivistiche dalla responsabilità dell'amministrazione attiva, degli uffici statali, amministrative, giudiziali, all'amministrazione archivistica”

3. **Termine ordinario dei 40 anni per i versamenti** `07:41` · _procedura_
   Secondo la normativa vigente, i versamenti dei documenti conservati negli archivi di deposito delle amministrazioni attive verso gli Archivi di Stato o l'Archivio Centrale dello Stato devono avvenire di norma 40 anni dopo l'esaurimento degli affari relativi.

   > “la legge prescite 40 anni dopo l'esaurimento degli affari. cioè prima i documenti stanno negli archivi di deposito”

6. **Segreto funzionale** `44:38` · _definizione_
   Il segreto funzionale è un tipo di segreto imposto dal presidente e dalla commissione parlamentare d'inchiesta sulla produzione documentaria, ma non è previsto da una norma di diritto positivo; è distinto dal segreto d'ufficio, che è disciplinato dal codice penale e da sanzioni amministrative.

   > “questo tipo di segreto che viene posto dal Presidente e dalla Commissione parlamentare d'inchiesta sulla documentazione prodotta è un segreto funzionale... non esiste una norma di diritto positivo che parli di questo segreto funzionale”
...

Real output, three of the eight points shown. 9 of the 69 points the model proposed were dropped because their quote was not in the recording.

Web demo

A small web interface (Gradio, Italian) is in demo/: upload a lecture or paste a link, choose how many points, download the notes. It runs locally, and is ready for Hugging Face Spaces:

pip install -r demo/requirements.txt
GEMINI_API_KEY=... GROQ_API_KEY=... python demo/app.py      # http://localhost:7860

Why another summarizer?

Most AI summarizers give you a paragraph you can't check. distillo is built for studying:

• You choose the density. --points 5 or --density low|medium|high (about 6, 12 or 24 points per hour). Every candidate point is ranked once, so the 5 points are always the top 5 of the 20. Changing the number reuses the cached ranking: it is instant and free.
• Every point is grounded. The model must quote the transcript word for word. distillo looks the quote up with fuzzy matching. If it isn't there, the point is dropped as a probable hallucination. If it is there, the quote tells us when the point was made, so the timestamp comes from the recording, not from the model's guess.
• Duplicates are merged. Teachers repeat the important things. Repetitions become one point, linked to the first time it was said.
• Free, or private. Gemini and Groq free tiers, with automatic fallback from one to the other when you hit a quota. Or --provider ollama --transcriber local to keep a recorded lesson on your machine.
• Notes in your language. A lecture in English, notes in Italian: --lang it.

Install

pipx install distillo              # or: uv tool install distillo
pipx install "distillo[local]"     # + offline transcription

From source:

git clone https://github.com/AndreaAcanfora/distillo && cd distillo
uv sync && uv run distillo --help

Set up a free model (two minutes)

distillo needs a language model to read the transcript. The free tiers are more than enough for lectures:

export GEMINI_API_KEY=...   # also read: GOOGLE_API_KEY, GOOGLE_GENERATIVE_AI_API_KEY
export GROQ_API_KEY=...

With both keys, --provider auto (the default) uses Gemini (gemini-3.6-flash) and moves to Groq (openai/gpt-oss-120b) when Gemini is overloaded, rate limited or out of quota. Groq's free tier allows 8,000 tokens per minute, so distillo sends it smaller pieces, one at a time: a 40-minute lecture takes about two minutes.

With ANTHROPIC_API_KEY set, Claude is added at the end of the chain. You can also pick a provider explicitly:

distillo talk.mp3 --provider anthropic                       # claude-opus-5-5
distillo talk.mp3 --provider anthropic -m claude-haiku-5-5   # cheaper
distillo talk.mp3 --provider ollama -m qwen3:8b              # local, nothing leaves your machine

Free tiers may use your prompts to improve the provider's products. For recordings that must stay private, use --provider ollama --transcriber local.

Usage

distillo lesson.mp4                                  # ~12 points per hour, in the spoken language
distillo lesson.mp4 -n 5                             # exactly 5 points
distillo lesson.mp4 --density high --lang en         # dense notes, in English
distillo lesson.mp4 -c "2nd year organic chemistry"  # context helps the model
distillo lesson.mp4 -o notes.md --save-transcript lesson.srt
distillo lesson.srt -n 20 --order rank               # reuse the transcript, most important first
distillo "https://www.youtube.com/watch?v=VIDEO_ID" -n 10   # a link: captions if any, else the audio
distillo captions.vtt --link "https://youtu.be/VIDEO_ID?t={seconds}"   # clickable timestamps
distillo lesson.mp4 -f json                          # for scripts and other apps

Inputs: any audio or video file, .srt / .vtt subtitles, plain .txt transcripts (no timestamps), or a link.

Links

YouTube and the other sites yt-dlp supports (Vimeo, archive.org, many university platforms, direct links to media files). If the video has captions in the spoken language, distillo uses them and skips transcription; otherwise it downloads only the audio. YouTube timestamps become links to the exact second. Use --no-captions to transcribe even when captions exist (automatic captions have no punctuation), and --cookies-from-browser firefox if YouTube asks you to sign in, which happens on servers and VPNs.

Speech to text

Groq accepts uploads up to 25 MB. Bigger recordings (a video, a WAV, a two-hour lecture) are converted to mono Opus at 24 kbit/s, about 11 MB per hour, and split into one-hour parts whose timestamps are joined back together.

How it works

audio/video ──Whisper──▶ timed transcript ──split──▶ parts of ~15 min
                                                         │ extract: every candidate point + verbatim quote
                                                         ▼
                                    grounding: find each quote in the transcript (fuzzy match)
                                       ├── not found → dropped (probable invention)
                                       └── found → timestamp of the matching segment
                                                         │ one ranking call over all points:
                                                         ▼ merge duplicates, order by importance
                                              top N ──▶ Markdown or JSON

Model responses are cached in ~/.cache/distillo (override with DISTILLO_CACHE_DIR), so trying a different number of points or another output format costs nothing. --no-cache forces fresh answers.

Roadmap

• v0.1: files and subtitles, Whisper (Groq or local), grounded points, density, merge, Markdown/JSON, free providers with fallback
• v0.2: links (YouTube and other sites, captions when available), automatic compression and splitting for Groq
• Web demo (Gradio, Italian interface)
• PyPI release
• Hosted demo on Hugging Face Spaces
• Chapters for long recordings
• Anki deck export and self-test questions
• --merge several lectures into one course revision sheet

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md: the test suite needs no API key.

Support

distillo is free and open source. If it saves you hours of re-watching lectures, you can buy me a coffee on PayPal. Thank you!

License

MIT © Andrea Acanfora