SenseAudio Meeting Summarizer

Transform meeting recordings into structured transcripts with speaker identification, timestamps, and local meeting summaries.

What This Skill Does

• Transcribe recorded meetings with official SenseAudio ASR endpoints
• Separate speakers with diarization on supported models
• Generate word and segment timestamps for navigation
• Produce local summaries and action-item candidates from transcripts
• Support optional realtime transcription for live meetings
• Export transcript and notes in text-friendly formats

Credential and Dependency Rules

• Read the API key from SENSEAUDIO_API_KEY.
• Send auth only as Authorization: Bearer <API_KEY>.
• Do not place API keys in query parameters, logs, or saved examples.
• If Python helpers are used, this skill expects python3, requests, and websockets.
• This skill should work without any external LLM credentials.
• If a user explicitly asks for LLM-based summarization, require a separately declared credential for that provider instead of assuming one exists.

Official ASR Constraints

Use the official SenseAudio ASR rules summarized below:

• HTTP endpoint: POST https://api.senseaudio.cn/v1/audio/transcriptions
• WebSocket endpoint: wss://api.senseaudio.cn/ws/v1/audio/transcriptions
• File upload limit: <=10MB per request
• Meeting-oriented HTTP model: sense-asr-pro
• Realtime WebSocket model: sense-asr-deepthink
• enable_speaker_diarization is supported only on sense-asr / sense-asr-pro
• max_speakers is documented only for sense-asr-pro
• enable_sentiment and timestamp_granularities[] are supported only on sense-asr / sense-asr-pro
• WebSocket audio must be pcm, 16000Hz, mono

Recommended Workflow

1. Validate the meeting asset:

• Prefer clear audio with limited background noise.
• Split files larger than 10MB before upload.

2. Transcribe with the right mode:

• Use HTTP sense-asr-pro for recorded meetings needing diarization and timestamps.
• Use WebSocket sense-asr-deepthink only for live streaming scenarios.

3. Request only needed features:

• For meeting notes, use response_format=verbose_json.
• Enable diarization, timestamps, and sentiment only when the user needs them.
• Provide max_speakers only when known and using sense-asr-pro.

4. Summarize locally first:

• Build summaries, decisions, and action-item candidates from the transcript itself.
• Keep the no-extra-credentials path as the default behavior.

5. Handle sensitive output carefully:

• Treat returned session_id, trace_id, and transcript contents as potentially sensitive.
• Do not expose provider identifiers unless needed for debugging.

Minimal HTTP Transcription Helper

import os

import requests

API_KEY = os.environ["SENSEAUDIO_API_KEY"]
API_URL = "https://api.senseaudio.cn/v1/audio/transcriptions"


def transcribe_meeting(audio_file, max_speakers=None, language=None, target_language=None):
    with open(audio_file, "rb") as handle:
        response = requests.post(
            API_URL,
            headers={"Authorization": f"Bearer {API_KEY}"},
            files={"file": handle},
            data={
                "model": "sense-asr-pro",
                "response_format": "verbose_json",
                "enable_speaker_diarization": "true",
                "enable_sentiment": "true",
                "enable_punctuation": "true",
                "timestamp_granularities[]": ["word", "segment"],
                **({"max_speakers": max_speakers} if max_speakers else {}),
                **({"language": language} if language else {}),
                **({"target_language": target_language} if target_language else {}),
            },
            timeout=300,
        )
    response.raise_for_status()
    return response.json()

Transcript Processing Pattern

• Read text for the full transcript
• Read segments for speaker-separated timeline entries
• Use speaker, start, end, text, and optional sentiment fields when present
• Use words only when word timestamps were requested

Local Summary Pattern

Generate notes from transcript structure without external services:

• summary: 3-6 bullets capturing the meeting arc
• decisions: statements containing agreements or final choices
• action_items: statements with owners, deadlines, or explicit follow-ups
• participants: derived from speaker labels
• timeline: ordered segments with timestamps

Heuristics that work without an LLM:

• Detect action items from patterns like will, need to, follow up, by Friday
• Detect decisions from patterns like decided, agreed, we will, final choice
• Aggregate speaker time by summing end - start

Realtime Meeting Pattern

For live meetings, use WebSocket only when streaming audio is actually available.

import asyncio
import json
import os

import websockets

API_KEY = os.environ["SENSEAUDIO_API_KEY"]
WS_URL = "wss://api.senseaudio.cn/ws/v1/audio/transcriptions"


async def transcribe_live_meeting(audio_stream):
    async with websockets.connect(
        WS_URL,
        additional_headers={"Authorization": f"Bearer {API_KEY}"},
    ) as ws:
        await ws.recv()
        await ws.send(json.dumps({
            "event": "task_start",
            "model": "sense-asr-deepthink",
            "audio_setting": {
                "sample_rate": 16000,
                "format": "pcm",
                "channel": 1,
            },
        }))

        async for audio_chunk in audio_stream:
            await ws.send(audio_chunk)

        await ws.send(json.dumps({"event": "task_finish"}))

Output Options

• Full transcript with timestamps in txt or json
• Meeting notes in md
• Action-item list in json or csv
• Speaker statistics in json
• Optional sentiment timeline when requested

Error Handling

• Poor audio quality: run an audio-quality check first or request a cleaner recording
• Large files: split recordings into chunks under 10MB
• Wrong language detection: set language explicitly on HTTP transcription
• Too many speakers: provide max_speakers only with sense-asr-pro
• Realtime failures: inspect WebSocket task_failed and base_resp.status_msg

Safety Notes

• Do not assume any external summarization provider credential exists.
• Do not add an LLM call unless the user asks for it and the skill metadata is updated to declare that credential.
• Prefer local transcript-based summarization for the default path.