Fal Whisper is an agent skill that connects to fal.ai’s Fast Whisper model to transcribe audio and video files into structured text. It is implemented in Python and uses an asynchronous queue — submit, poll, retrieve — to handle transcription jobs, which makes it more stable than synchronous approaches when processing large files.
The skill accepts both local files and remote files hosted at public URLs. Local files are sent via base64 encoding; remote files are referenced by URL. Supported formats include MP3, MP4, M4A, WAV, and FLAC, among others.
Transcription results are stored locally in a dedicated directory (~/.fal-whisper/). Each job produces at minimum a plain .txt file containing the full transcript. Optionally, the skill generates a SubRip (.srt) subtitle file with precision timestamps, suitable for import into video editing applications such as Premiere Pro or DaVinci Resolve. Word-level timestamps are also available.
Because audio transcription is computationally intensive, offloading it to fal.ai’s hardware avoids the overhead of running inference locally. Structured output — including text chunks and segment timestamps — makes downstream processing straightforward for agents that need to search, index, or caption audio content.
This skill is not an MCP server and exposes no callable tools in the MCP sense; it is a Python-based skill for agent workflows that already integrate with the fal.ai platform.