Audio/Video Transcription & Summarization#

Transcribe audio/video files using the SenseASR API ( api.senseaudio.cn), then summarize the content into structured notes.

{baseDir} refers to this skill's directory.

Prerequisites#

• Environment variable SENSEAUDIO_API_KEY configured (get your key at https://senseaudio.cn/platform/api-key)
• Python 3.8+ with requests installed
• For large files (>10MB): ffmpeg installed for splitting（macOS: brew install ffmpeg，Windows: ffmpeg.org 下载并加入 PATH，Linux: apt install ffmpeg）

Quick Start#

1. Run the transcription script:

bash

python {baseDir}/scripts/transcribe.py <audio_file> [--model sense-asr-pro] [--language zh] [--speakers] [--sentiment] [--translate en]

1. The script outputs a transcript .txt file alongside the source file
2. Read the transcript and generate a summary (see Summary Format below)

Workflow#

Step 1: Assess the Audio File#

Check file size and format:

• Supported formats: wav, mp3, ogg, flac, aac, m4a, mp4
• Max file size per request: 10MB
• If file > 10MB, the script auto-splits using ffmpeg

Step 2: Choose the Right Model#

ModelUse Whensense-asr-liteQuick batch transcription, simple audio, cost-sensitivesense-asrGeneral transcription, need speaker separation or timestampssense-asr-proHigh accuracy needed: meetings, interviews, complex audiosense-asr-deepthinkNoisy audio, dialects, heavy jargon, speech-to-clean-text

Default to sense-asr-pro for best quality.

Step 3: Transcribe#

Run the transcription script. Key options:

bash

# Basic transcription
python {baseDir}/scripts/transcribe.py recording.mp3

# Meeting with multiple speakers + emotion
python {baseDir}/scripts/transcribe.py meeting.wav \
  --model sense-asr-pro \
  --speakers --max-speakers 4 \
  --sentiment \
  --timestamps segment

# Transcribe and translate to English
python {baseDir}/scripts/transcribe.py lecture.mp3 \
  --model sense-asr \
  --translate en

Step 4: Summarize#

After transcription, read the transcript file and produce a summary using the format below.

Summary Format#

Generate summaries in this structure:

markdown

# [Title - inferred from content]

**Source**: filename.mp3
**Duration**: X min Y sec
**Date**: YYYY-MM-DD
**Speakers**: [if speaker diarization was used]

## Key Points
- Point 1
- Point 2
- ...

## Detailed Summary
[2-4 paragraph summary of the content organized by topic/chronology]

## Action Items
- [ ] Action item 1 (assigned to Speaker X, if applicable)
- [ ] Action item 2

## Notable Quotes
> "Direct quote from transcript" — Speaker X, [timestamp if available]

## Full Transcript
<details>
<summary>Click to expand full transcript</summary>

[Full transcript text here, with speaker labels and timestamps if available]

</details>

Adapt the template based on content type:

• Meeting: emphasize action items, decisions, speaker contributions
• Lecture/Talk: emphasize key concepts, learning points, structure
• Interview: emphasize Q&A pairs, key responses
• Podcast: emphasize topics discussed, interesting insights

API Reference#

For full SenseASR API parameters and response formats, see api-reference.md.