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

agent-media 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)

Output

Returns a JSON object with transcription data:

{
  "ok": true,
  "media_type": "audio",
  "action": "transcribe",
  "provider": "fal",
  "output_path": "transcription_123_abc.json",
  "transcription": {
    "text": "Full transcription text...",
    "language": "en",
    "segments": [
      { "start": 0.0, "end": 2.5, "text": "Hello.", "speaker": "SPEAKER_0" },
      { "start": 2.5, "end": 5.0, "text": "Hi there.", "speaker": "SPEAKER_1" }
    ]
  }
}

Examples

Basic transcription (auto-detect language):

agent-media audio transcribe --in interview.mp3

Transcription with speaker identification:

agent-media audio transcribe --in meeting.wav --diarize

Transcription with specific language and speaker count:

agent-media audio transcribe --in podcast.mp3 --diarize --language en --speakers 3

Use specific provider:

agent-media audio transcribe --in audio.wav --provider replicate

Extracting Audio from Video

To transcribe a video file, first extract the audio:

# Step 1: Extract audio from video
agent-media audio extract --in video.mp4 --format mp3

# Step 2: Transcribe the extracted audio
agent-media audio transcribe --in extracted_xxx.mp3

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

agent-media audio transcribe --in audio.mp3 --provider local

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