Vibe Transcribe

Performs speech-to-text transcription on audio and video files using local machine learning models (likely Whisper or similar) that run entirely on-device without cloud API calls. The system handles multiple audio formats and video containers, extracting audio streams and processing them through a local inference pipeline that maintains privacy and eliminates per-minute API costs.

Automatically detects and extracts audio streams from diverse video container formats (MP4, MKV, WebM, etc.) and normalizes audio to a standard format for downstream transcription processing. Uses container-aware parsing (likely FFmpeg or libav) to handle codec detection, stream selection, and format conversion without manual user configuration.

Exposes transcription functionality via HTTP REST API, allowing external applications to submit files for transcription and retrieve results. Supports asynchronous job submission, polling for status, and webhook callbacks for result notification. Likely uses a lightweight HTTP framework (Flask, FastAPI) with job queue integration.

Processes multiple audio/video files sequentially or in parallel with real-time progress reporting, queue management, and error handling. Tracks transcription status per file, allows pause/resume, and provides detailed logs of successes and failures without requiring manual orchestration or external job queue systems.

Generates transcriptions with precise word-level or sentence-level timestamps, supporting multiple output formats (SRT, VTT, JSON) for subtitle generation and media synchronization. Preserves timing information from the speech model's output and formats it according to standard subtitle specifications or custom JSON schemas.

Automatically detects the spoken language in audio and selects the appropriate transcription model or language-specific parameters. Supports transcription of multiple languages without requiring users to manually specify language codes, with fallback handling for mixed-language content.

Identifies and separates different speakers in audio, attributing transcribed segments to specific speakers with labels (Speaker 1, Speaker 2, etc.). Uses voice activity detection and speaker embedding models to cluster and distinguish speakers without requiring speaker enrollment or training data.

Provides a browser-based interface allowing users to drag-and-drop audio/video files for transcription without command-line interaction. The UI handles file upload, progress visualization, and result display, with optional export options. Likely runs a local HTTP server that processes files and streams results back to the browser.

Allows users to choose between different model sizes (tiny, base, small, medium, large) and configure transcription parameters like language, temperature, and beam search settings. Exposes model-specific options without requiring code changes, enabling trade-offs between speed, accuracy, and resource usage.

Exports transcription results in multiple formats (plain text, SRT, VTT, JSON, Markdown) with customizable formatting and metadata inclusion. Supports batch export of multiple files and template-based formatting for custom output structures.

Automatically detects GPU availability (CUDA, Metal, ROCm) and uses GPU acceleration when available, with transparent fallback to CPU processing if GPU is unavailable or incompatible. Handles device memory management and batch sizing to prevent out-of-memory errors.