Audio Transcription

Reusable local audio transcription via faster-whisper. Hebrew-first (uses the ivrit-ai fine-tune by default), but works for any language. Point it at a recording, get a timestamped Markdown transcript. Nothing leaves the machine.

When to use

• The user has a recording (phone call, Google Meet, voice memo, interview, lecture, podcast) and wants it transcribed.
• The user asks to "תמלל", "תמלול", "transcribe", "convert speech to text".
• A workflow needs to read what was said in audio/video before proceeding (e.g. drafting a document from a kickoff call, extracting decisions from a meeting).

When NOT to use

• The recording is already transcribed (.txt / .srt / .vtt / .md beside it) — read that instead.
• The user asks for diarization (who-said-what speaker labels). This skill does not do diarization out of the box (see "Future work" below). Acknowledge the limitation and offer either: timestamped transcript + context-based attribution, or a separate pyannote.audio setup.
• The audio is on a remote service requiring API auth (YouTube link, Zoom cloud, etc.) — first get a local file.

Setup

pip install faster-whisper
# plus ffmpeg from your system package manager - see ../install-whisper/

Quickstart

# Hebrew (default model: ivrit-ai/whisper-large-v3-turbo-ct2)
python transcribe.py "<path-to-audio>" -l he -o transcripts\<name>.md

# English / auto-detect
python transcribe.py "<path-to-audio>" -o transcripts\<name>.md

The script is also invokable from bash:

python transcribe.py "<path-to-audio>" -l he -o transcripts/<name>.md

Defaults & flags

Output format

Markdown with one segment per paragraph:

# Transcript: <filename>

- Source: ...
- Model: ...
- Language: he (99%) (forced)
- Duration: 0:14:36
- VAD: Silero (on)
- Beam size: 5

---

<div dir="rtl">

`00:00:00` שלום, אז דיברנו על...

`00:00:08` כן, אני חושב שצריך לבדוק...

</div>

Hebrew transcripts are automatically wrapped in <div dir="rtl"> for correct RTL rendering on GitHub and in Markdown previewers.

Recipe: project-local placement

When invoked inside a project, save the transcript under the project's working tree, not next to the source file in Downloads. Pick the most specific subdirectory available:

1. If the recording is conference/event-related → conference/transcripts/
2. If it's interview source material for a review → sources/transcripts/
3. Generic working call → transcripts/

Filename convention: <YYYY-MM-DD>-<short-slug>.md. Example: 2026-05-27-team-call.md.

Performance expectations

On CPU (no GPU), expect roughly:

• large-v3 / ivrit-ai/...turbo-ct2: 0.3-1.4× realtime depending on CPU and how much silence the VAD skips
• medium: 1-2× realtime
• small: 4-6× realtime

The first run downloads the model (~1.5-3 GB) into the HuggingFace cache. Subsequent runs are warm and fully offline.

If you have an NVIDIA GPU, --device cuda --compute-type float16 is 5-20× faster than CPU.

Hebrew quality notes

• ivrit-ai/whisper-large-v3-turbo-ct2 is the state-of-the-art Hebrew Whisper fine-tune. It significantly outperforms vanilla large-v3 on Hebrew WER, especially for spontaneous speech and accented Hebrew.
• For mixed Hebrew/English calls, the ivrit-ai model handles code-switching well; English-only segments will transcribe correctly too.
• Numbers/dates often come out as words; let the user polish if they need digits.
• Phone-line audio (8 kHz) is noticeably worse than meeting-quality audio. Consider --beam-size 8 and revisit segments with low semantic coherence.

Future work (not implemented)

• Speaker diarization (who-said-what): would require pyannote.audio + PyTorch + a HF-gated model accept. For two-speaker calls, an LLM can often re-attribute speakers post-hoc from context.
• SRT/VTT output: easy to add — currently only Markdown is emitted.
• Force-alignment / phrase-level confidence: faster-whisper exposes word_timestamps and per-segment probabilities; not surfaced in MD yet.

Privacy

The transcription is fully local — audio never leaves the machine. Models are downloaded once from HuggingFace; runtime is offline.