长音频 ASR 工作流

# Long Audio ASR Workflow

## When to use

Prepare long audio or video for speech recognition and return timestamped, speaker-aware transcripts with uncertainty notes.

## Workflow

1. Confirm the user owns the input or has permission to process it. Collect the source material, audience, output format, and acceptance criteria.
2. Separate observed facts and user-provided constraints from inference. Ask for missing information instead of inventing it.
3. Inspect duration, language, channel quality, speaker count, timing needs, and privacy constraints before transcription or subtitle work.
4. Use <skill-tag>skill:820e8b64-9a9c-5c33-8337-ddd348f77f3a</skill-tag> for the connected media workflow. Return timestamps and mark uncertain words rather than guessing.

## Output contract

Return the input assumptions, the approved plan, the generated or analyzed result, uncertainty and limitations, and a concise verification checklist. Preserve source links and provider-returned identifiers when available. Stop if the required connected capability is unavailable; do not substitute local scripts, package installation, or embedded credentials.

## Safety and review

Do not upload private or sensitive media without explicit authorization. Do not claim certainty beyond the evidence. Require review before externally visible, paid, destructive, or account-changing actions.

## Source and license

This Aident-native re-authoring is based on [the pinned upstream source](https://github.com/bytedance/agentkit-samples/tree/cc8fd09a63fe8b07c897e02c2b695fe4476fff04/skills/byted-las-asr-pro), discovered through [ModelScope](https://modelscope.cn/skills/@bytedance/byted-las-asr-pro), and used under Apache-2.0. It preserves the upstream workflow goal while removing local executable, credential, and package assumptions.