Volcengine ASR（lecture-to-md 子 skill）

Use the deterministic scripts in scripts/; do not reproduce API requests manually.

本 skill 是 lecture-to-md 的云端 ASR 子 skill。日常请走 local-asr/（sherpa-onnx X-ASR，免费、跨平台、隐私友好）；只有当用户明确要求云端准确率，或本地模型不足以应对时，才启用本 skill。

Choose a mode

• For turbo-v1, run scripts/turbo_transcribe.py. It accepts a local audio/video file, extracts a 16 kHz mono MP3, sends Base64 in one synchronous request, and writes JSON/TXT/SRT/Markdown.
• For standard-v1, idle-v1, or standard-v2, first run scripts/upload_submit.py, then run scripts/poll_transcript.py with the resulting job.json.
• Prefer standard-v2 when the user has opened 豆包录音文件识别模型 2.0.
• Local asynchronous inputs require TOS credentials. If a direct-download HTTP(S) URL already exists, pass --audio-url and skip TOS.

Read references/api-modes.md before changing endpoints, resource IDs, limits, or request fields.

Credentials

Read credentials from .secret in this skill directory. Never print or copy credential values into results. .secret must remain gitignored and mode 0600.

The ASR v3 calls use only VOLCENGINE_ASR_APP_ID and VOLCENGINE_ASR_ACCESS_TOKEN. VOLCENGINE_ASR_SECRET_KEY is retained locally but not sent. TOS uses a separate account-level AK/SK, endpoint, region, and bucket.

Commands

When .venv is absent, prepare the optional TOS dependency once:

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt

upload_submit.py automatically re-executes with this environment for local TOS uploads.

Turbo 1.0:

python3 scripts/turbo_transcribe.py "/absolute/path/video.mp4"

Submit a local 2.0 standard job through TOS:

python3 scripts/upload_submit.py \
  --mode standard-v2 \
  --input "/absolute/path/video.mp4"

Poll the generated job:

python3 scripts/poll_transcript.py "/absolute/path/job.json"

For a pre-existing direct URL:

python3 scripts/upload_submit.py \
  --mode standard-v2 \
  --audio-url "https://example.com/audio.mp3"

Operating rules

• Check that the requested service is open before a paid call. Do not invent undocumented seedasr Turbo or idle resource IDs.
• When testing multiple modes, use different source videos so paid/free ASR time produces useful, non-duplicate transcripts.
• Keep the TOS bucket private and use only short-lived pre-signed GET URLs.
• Leave a reusable job.json when polling times out. Delete the TOS object only after a terminal success/failure unless the user passes --keep-upload.
• Report the absolute paths of transcript.txt, transcript.srt, and result.json after success.