Speech to Text

Transcribe or translate audio files to text using a public Hugging Face Whisper Space over Gradio. Use when the user sends voice notes, audio attachments, me...

Security Scan

The skill is internally coherent: it transcribes local audio by uploading it to a public Hugging Face Gradio Whisper Space, requests no secrets, and its files and instructions match its stated purpose.

Purpose & Capability

Name/description claim using a public Hugging Face Whisper Space; the included script and SKILL.md both implement exactly that (upload to a Gradio Space, call predict, wait for result). No unrelated binaries, env vars, or services are requested.

Instruction Scope

Instructions explicitly tell the agent to read a local audio file and upload it to a public Gradio/Hugging Face Space. This is expected for the stated purpose but has privacy implications (documented in guardrails). The skill does not attempt to read other files or arbitrary system state.

Install Mechanism

No install spec; skill is instruction + a small Python script. No external downloads or package installs are performed by the skill itself.

Credentials

No credentials or sensitive environment variables are required. The only optional environment variable (HF_WHISPER_SPACE) is used to override the target space URL and is justified by the purpose.

Persistence & Privilege

Skill is not declared always:true and does not request persistent system privileges. It runs as an on-demand script and does not modify other skills or global agent settings.

Assessment

This skill works as advertised: it reads a local audio file and uploads it to a public Hugging Face Gradio Space for Whisper-based transcription. Before installing or using it, consider privacy and trust: the default endpoint is a third-party public space (hf-audio-whisper-large-v3-turbo.hf.space), so do not send highly sensitive audio unless you accept third-party processing. You can override the space with HF_WHISPER_SPACE or --space to point to a self-hosted or trusted endpoint. Verify the space URL you use is trustworthy, and be aware of rate limits, queueing, and potential outages. The script makes outbound HTTP requests and prints results; review or audit the target space if you need confidentiality guarantees.

Use this skill to turn local audio files into text with a public Whisper-based endpoint.

Quick start

Run:

python3 scripts/transcribe.py /path/to/file.ogg

Return the transcript as plain text. By default, the script also applies lightweight Chinese punctuation and sentence-breaking cleanup.

For machine-readable output:

python3 scripts/transcribe.py /path/to/file.ogg --json

To disable cleanup and keep the raw model text:

python3 scripts/transcribe.py /path/to/file.ogg --format raw

To force Chinese punctuation cleanup:

python3 scripts/transcribe.py /path/to/file.ogg --format zh

For English translation instead of same-language transcription:

python3 scripts/transcribe.py /path/to/file.ogg --task translate

Workflow

1. Confirm the input is a local audio file.
2. Run scripts/transcribe.py on it.
3. If the transcript looks imperfect, tell the user it came from a public Whisper endpoint and may need cleanup.
4. If helpful, post-process into:
   • cleaned transcript
   • summary
   • action items
   • bilingual output

What the script does

The script:

• uploads the local file to a public Gradio-backed Hugging Face Space
• submits a Whisper transcription job
• waits for completion via the Gradio event stream
• prints the resulting text

Default endpoint:

• https://hf-audio-whisper-large-v3-turbo.hf.space

Override it with:

python3 scripts/transcribe.py input.ogg --space https://your-space.hf.space

or set:

export HF_WHISPER_SPACE=https://your-space.hf.space

Guardrails

• Treat this as a best-effort public/free path, not a privacy-grade path.
• Do not use for highly sensitive audio unless the user explicitly accepts public third-party processing.
• Expect rate limits, queueing, and occasional outages.
• If the public endpoint fails, explain that the free backend is unavailable and offer alternatives.

Output handling

Prefer to return:

• the raw transcript when the user asked to "转文字/听写"
• a cleaned version when punctuation is poor
• a short note about uncertainty if names, numbers, or jargon may be wrong

Script

• scripts/transcribe.py — public Whisper transcription helper