kouhxp/yapsnap

What it solves

yapsnap is a command-line tool that converts video URLs or audio files into plaintext transcripts. It eliminates the need for expensive GPUs, cloud APIs, or paid subscriptions by performing all transcription, language detection, and speaker diarization locally on the user's CPU.

How it works

The tool follows a five-step pipeline:

1. Fetch: Uses yt-dlp to download audio from URLs (YouTube, TikTok, X, Instagram, etc.) or accesses local files.
2. Identify: Automatically detects the spoken language using a pruned whisper-tiny language-ID head (INT8 ONNX) by sampling small windows of the audio.
3. Decode: Uses ffmpeg to convert media into 16 kHz mono PCM audio.
4. Recognize: Employs a streaming Zipformer2 transducer (Kroko) via the sherpa-onnx runtime to transcribe audio into text on the CPU.
5. Format: Outputs the result as a single paragraph or as sentence-level timestamps. If --diarize is enabled, it runs a second pass with speaker-segmentation and embedding models to label who spoke when.

Who it’s for

It is designed for users who need fast, private, and offline transcription of web videos or local recordings without requiring specialized hardware (CUDA/M-series) or API keys.

Highlights

• CPU-Only Performance: Optimized for plain CPU cores, often transcribing faster than real-time.
• Broad Source Support: Handles a wide array of social media URLs and local audio/video formats via yt-dlp and ffmpeg.
• Privacy-First: Models are cached locally after the first run; audio never leaves the machine.
• Automatic Language Detection: Supports ten languages (English, German, Spanish, French, Italian, Hebrew, Dutch, Portuguese, Swedish, Turkish) with auto-detection.
• Speaker Diarization: Can identify and label different speakers (SPEAKER_00, SPEAKER_01) using ONNX models.
• Lean Dependencies: Avoids heavy frameworks like PyTorch, relying instead on sherpa-onnx, numpy, and yt-dlp.