audio-transcribe skill by agntswrm

Audio Transcribe

Transcribes audio files to text with timestamps. Supports automatic language detection, speaker identification (diarization), and outputs structured JSON with segment-level timing.

Command

npx agent-media@latest audio transcribe --in <path> [options]

Inputs

Option	Required	Description
`--in`	Yes	Input audio file path or URL (supports mp3, wav, m4a, ogg)
`--diarize`	No	Enable speaker identification
`--language`	No	Language code (auto-detected if not provided)
`--speakers`	No	Number of speakers hint for diarization
`--out`	No	Output path, filename or directory (default: ./)
`--provider`	No	Provider to use (local, fal, replicate, runpod)

Output

Returns a JSON object with transcription data:

Examples

Basic transcription (auto-detect language):

Transcription with speaker identification:

Transcription with specific language and speaker count:

Use specific provider:

Extracting Audio from Video

To transcribe a video file, first extract the audio:

Providers

local

Runs locally on CPU using Transformers.js, no API key required.

Uses Moonshine model (5x faster than Whisper)
Models downloaded on first use (~100MB)
Does NOT support diarization — use fal or replicate for speaker identification
You may see a mutex lock failed error — ignore it, the output is correct if "ok": true

fal

Requires FAL_API_KEY
Uses wizper model for fast transcription (2x faster) when diarization is disabled
Uses whisper model when diarization is enabled (native support)

replicate

Requires REPLICATE_API_TOKEN
Uses whisper-diarization model with Whisper Large V3 Turbo
Native diarization support with word-level timestamps

runpod

Requires RUNPOD_API_KEY
Uses pruna/whisper-v3-large model (Whisper Large V3)
Does NOT support diarization (speaker identification) - use fal or replicate for diarization