Whisper Api Stt

The problem

Large audio files often exceed API limits or consume excessive bandwidth, while local transcription models lack the accuracy of specialized cloud endpoints. Managing multipart form data and parsing complex JSON segments manually is tedious and error-prone.

What it does

• Pre-compresses audio to mono 16 kHz Opus/MP3 using ffmpeg to fit under 25 MB limits.
• Executes authenticated multipart uploads to OpenAI transcription endpoints using curl.
• Parses verbose_json responses into human-readable formats with jq.
• Implements diarization models to extract speaker labels from recordings.
• Optimizes accuracy for technical jargon and names using prompt-injection hints.

Frameworks & tools

OpenAI API, ffmpeg, curl, jq, and standard Unix shell environments.

Why this beats prompting it yourself

This skill handles the low-level infrastructure of audio processing that standard LLM prompts ignore. It ensures files are correctly downsampled for cost-efficiency and provides the specific jq filters needed to turn raw API responses into usable transcripts immediately.

Use cases

• Transcribe long podcasts or meetings by compressing them for API compatibility.
• Generate timestamped SRT or VTT files for video subtitles.
• Identify different speakers in interview recordings using diarization models.
• Improve transcription of technical documentation by biasing the model with specific jargon.

Known limitations

Requires an active OPENAI_API_KEY and local installations of ffmpeg and jq. Files that remain over 25 MB after compression must be manually chunked.