digituva
Claude Code plugins by Digituva Yazılım
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/digituvacom--claude-listen-audio)HTML
<a href="/listings/digituvacom--claude-listen-audio"><img src="/api/listings/digituvacom--claude-listen-audio/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
As a plugin (recommended)
Inside Claude Code, run:
/plugin install listen-audio --marketplace digituvacom/claude-listen-audio
<sub>On Claude Code older than 2.1.275, use two steps: `/plugin marketplace add digituvacom/claude-listen-audio`, then `/plugin install listen-audio@digituva`.</sub>
Then install the Python dependencies once (Python 3.9+):
python -m pip install faster-whisper librosa matplotlib soundfile
Optional, for sound-event tagging:
python -m pip install transformers torch
Updates arrive with `/plugin update listen-audio`. If a dependency is missing, Claude notices on first use and tells you what to install.
As a plain skill
git clone https://github.com/digituvacom/claude-listen-audio ~/.claude/skills/listen-audio python -m pip install -r ~/.claude/skills/listen-audio/requirements.txt
<sub>Windows: `%USERPROFILE%\.claude\skills\listen-audio`. Restart Claude Code afterwards.</sub>
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 17 lines
- {
- "name": "digituva",
- "description": "Claude Code plugins by Digituva Yazılım",
- "owner": {
- "name": "Digituva Yazılım",
- "url": "https://github.com/digituvacom"
- },
- "plugins": [
- {
- "name": "listen-audio",
- "source": "./",
- "description": "Give Claude ears: Whisper transcription, acoustic analysis, sound-event tags and spectrograms for any audio or video file.",
- "category": "productivity",
- "tags": ["audio", "speech-to-text", "whisper", "spectrogram"]
- }
- ]
- }
README.md · README · 232 lines
- <div align="center">
- # 🎧 claude-listen-audio
- ### Give Claude ears.
- A [Claude Code](https://claude.com/claude-code) skill that turns any recording into something Claude can understand:<br>
- a timestamped transcript, precise acoustic measurements, and a spectrogram it can actually *look at*.
- [](https://claude.com/claude-code)
- [](https://www.python.org/)
- [](https://github.com/SYSTRAN/faster-whisper)
- [](LICENSE)
- **English** · [Türkçe](README.tr.md)
- <a href="https://www.youtube.com/watch?v=knZadfUNXOc"><img src="https://img.youtube.com/vi/knZadfUNXOc/maxresdefault.jpg" alt="Watch the 34-second listen-audio demo on YouTube" width="860"></a>
- <sub>▶ <a href="https://www.youtube.com/watch?v=knZadfUNXOc">Watch the 34-second demo</a> · <a href="https://www.youtube.com/watch?v=XZR8gpV7_Ys">Türkçe video</a></sub>
- <br><br>
- <img src="docs/example-spectrogram.png" alt="Waveform, loudness curve and mel spectrogram of a test clip: speech, silence, a 440 Hz tone and white noise" width="860">
- <sub>One clip, four sounds: speech (0–5.7 s) · silence · a pure 440 Hz tone · white noise. Claude reads this image and tells them apart.</sub>
- </div>
- ---
- ## Contents
- - [Why](#why)
- - [What Claude gets](#what-claude-gets)
- - [How it works](#how-it-works)
- - [Install](#install)
- - [Usage](#usage)
- - [Example output](#example-output)
- - [Performance](#performance)
- - [Limitations](#limitations)
- - [FAQ](#faq)
- - [License](#license)
- ## Why
- Claude reads text and sees images, but it **cannot hear**. Hand it a voice note, a meeting recording or a song and it has nothing to work with.
- This skill closes that gap. It translates sound into the two forms Claude handles best, **numbers** and **pictures**, so it can answer questions like:
- > *"What does this voice note say?"* · *"Summarize this 40-minute meeting."* · *"Is there background noise or hum?"*<br>
- > *"What's the tempo and key of this track?"* · *"When does the music start?"* · *"Is the audio clipping?"*
- ## What Claude gets
- | | Capability | Details |
- |:-:|---|---|
- | 🗣️ | **Transcript** | Timestamped speech-to-text via faster-whisper. Auto-detects 99 languages, Turkish included. Low-confidence lines are flagged, not passed off as fact. |
- | 📊 | **Acoustic report** | Loudness (peak / RMS / dynamic range), clipping, silences and pauses, pitch and musical note, tempo (BPM), key, timbre, energy per frequency band. |
- | 🕒 | **Timeline** | How the sound evolves window by window: level, brightness, noisiness, onsets, dominant pitch class. |
- | 🖼️ | **Spectrogram** | Waveform + loudness curve + mel spectrogram as one PNG. Claude opens it and *sees* speech harmonics, tones, beats and noise. |
- | 🏷️ | **Sound events** *(optional)* | AudioSet labels per 10 s chunk: Dog, Siren, Applause, Guitar, Car horn… 527 classes. |
- Works with **mp3, wav, m4a, ogg, flac, opus, webm, mp4** and most other audio/video containers. No system ffmpeg required.
- ## How it works
- ```mermaid
- flowchart LR
- A[🎵 Audio / video file] --> B[Decode<br/>PyAV]
- B --> C[faster-whisper]
- B --> D[librosa analysis]
- B --> E[AudioSet classifier<br/><i>optional</i>]
- C --> F[📝 transcript]
- D --> G[📊 metrics + timeline]
- D --> H[🖼️ spectrogram.png]
- E --> I[🏷️ sound tags]
- F & G & H & I --> J((Claude))
- ```
- Everything runs **locally** on your machine. No audio is uploaded anywhere; only the model weights are downloaded once from Hugging Face.
- ## Install
- ### As a plugin (recommended)
- Inside Claude Code, run:
- ```
- /plugin install listen-audio --marketplace digituvacom/claude-listen-audio
- ```
- <sub>On Claude Code older than 2.1.275, use two steps: `/plugin marketplace add digituvacom/claude-listen-audio`, then `/plugin install listen-audio@digituva`.</sub>
- Then install the Python dependencies once (Python 3.9+):
- ```bash
- python -m pip install faster-whisper librosa matplotlib soundfile
- ```
- Optional, for sound-event tagging:
- ```bash
- python -m pip install transformers torch
- ```
- Updates arrive with `/plugin update listen-audio`. If a dependency is missing, Claude notices on first use and tells you what to install.
- ### As a plain skill
- ```bash
- git clone https://github.com/digituvacom/claude-listen-audio ~/.claude/skills/listen-audio
- python -m pip install -r ~/.claude/skills/listen-audio/requirements.txt
- ```
- <sub>Windows: `%USERPROFILE%\.claude\skills\listen-audio`. Restart Claude Code afterwards.</sub>
- ## Usage
- ### With Claude
- Just mention a file. The skill triggers on its own:
- ```
- > What does this say? ~/Downloads/voice-note.m4a
- > Summarize the decisions in meeting.mp3
- > Is there a 50 Hz hum in this recording?
- > What's the BPM and key of track.wav?
- ```
- ### From the command line
- ```bash
- python scripts/listen.py recording.mp3 --out ./out
- ```
- | Flag | Default | Description |
- |---|---|---|
- | `--mode` | `all` | `transcribe`, `analyze` or `all` |
- | `--model` | `small` | Whisper size: `tiny` · `base` · `small` · `medium` · `large-v3` · `turbo` |
- | `--language` | auto | Force a language code, e.g. `en`, `tr`, `de` |
- | `--tags` | off | Add AudioSet sound-event labels |
- | `--start` / `--end` | full file | Analyse only a time range (seconds) |
- | `--device` | `auto` | `cuda` or `cpu`; falls back to CPU automatically |
- | `--timeline` | `12` | Number of timeline windows |
- **Outputs** written to `--out`:
- ```
- out/
- ├── report.md # human-readable report (also printed to stdout)
- ├── report.json # every number, machine-readable
- ├── transcript.txt # [mm:ss → mm:ss] lines
- └── spectrogram.png # the picture Claude looks at
- ```
- ## Example output
- <details open>
- <summary><b>report.md</b> for the test clip shown above</summary>
- ```markdown
- ## Speech
- language **en** (p=0.996), 5.8 s of speech, 1 segments, 0 low-confidence
- [00:00.00 → 00:05.76] Hello, this is an audio test. The meeting will start at 3 o'clock this afternoon.
- ## Acoustics
- - Loudness: peak 0.0 dBFS, RMS -15.2 dBFS, dynamic range 21.3 dB, clipping 0.031%
- - Silence: 35% of the time, 5 sound regions; gaps ≥1s: 00:05.67–00:08.31
- - Pitch: voiced 50%, median 194.9 Hz (G3), range D3–A4
- ### Timeline
- | from | to | RMS dBFS | centroid Hz | flatness | pitch class |
- |----------|----------|----------|-------------|----------|-------------|
- | 00:08.24 | 00:09.27 | -11.5 | 442 | 0.0 | A | ← the 440 Hz tone
- | 00:11.33 | 00:12.36 | -9.8 | 5667 | 0.548 | D# | ← white noise
- ```
- </details>
- ## Performance
- Rough transcription speed with the default `small` model:
- | Hardware | Speed |
- |---|---|
- | NVIDIA GPU (CUDA 12) | ~20–40× real time |
- | Modern laptop CPU | ~1–2× real time |
- First use of each model downloads it once: `base` ≈ 150 MB · `small` ≈ 500 MB · `turbo` ≈ 1.6 GB.
- The optional tagger downloads ≈ 350 MB.
- ## Limitations
- - **No speaker diarization.** Whisper doesn't separate voices; "who said what" is inferred from content and pauses and labelled as such.
- - **Music metrics on speech are noise.** Tempo and key are only meaningful for music.
- - **Voice hints are hints.** Gender/age guesses come from median pitch alone and can be wrong.
- - **Pitch analysis covers the first 90 s** to keep long files fast.
- ## FAQ
- <details>
- <summary><b>Does my audio leave my computer?</b></summary>
- No. Decoding, transcription and analysis all run locally. The only network traffic is the one-time model download from Hugging Face.
- </details>
- <details>
- <summary><b>The GPU isn't used.</b></summary>
- faster-whisper needs the CUDA 12 cuBLAS and cuDNN libraries. Without them the script prints a notice and continues on CPU, so nothing breaks.
- </details>
- <details>
- <summary><b>The transcript is wrong in places.</b></summary>
- Use a larger model (`--model turbo` or `large-v3`) and, for short or mixed-language clips, force the language with `--language`.
- </details>
- <details>
- <summary><b>It's a 2-hour file.</b></summary>
- Use `--start` / `--end` for a first pass, or run the whole thing in the background. Claude will do this automatically for long files.
- </details>
- ## License
- [MIT](LICENSE) © 2026 Digituva Yazılım
- <div align="center">
- <sub>Built with Digituva Yazılım · Powered by <a href="https://github.com/SYSTRAN/faster-whisper">faster-whisper</a>, <a href="https://librosa.org">librosa</a> and <a href="https://matplotlib.org">matplotlib</a></sub>
- </div>
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.