AI Agent Hub
Back to skills
TencentCloud ASR icon

TencentCloud ASR

Design & Media Updated 2026.08.30

Paste the following prompt into your AI chat to install this skill:

Please install @tencent-adm/tencentcloud-asr according to the official guide at https://skillhub.cn/install/skillhub.md.

About this skill

Specific Problem

In software development, speech-to-text is a common requirement for tasks like meeting transcription, voice interaction, or content analysis. Traditional solutions often face challenges such as low accuracy, limited multilingual support, poor real-time performance, or high costs. TencentCloud ASR addresses these pain points by providing a high-accuracy, low-latency speech-to-text service, covering scenarios from short sentences to long recordings.

Core Capabilities and Key Steps

TencentCloud ASR offers multiple recognition modes, selected based on audio characteristics:

  • Sentence Recognition for short audio (≤60s, ≤3MB), using the sentence_recognize.py script for quick results.
  • Flash Recognition for medium audio (≤2h, ≤100MB), using flash_recognize.py to balance speed and resources.
  • File Recognition for large files or public URLs, using file_recognize.py for asynchronous processing.

Key steps include:

  1. Audio Inspection: Run inspect_audio.py first to analyze audio format, duration, size, etc., to inform subsequent routing.
  2. Credential Check: Use self_check.py to validate TencentCloud credentials like SecretId and SecretKey, ensuring service availability.
  3. Engine Selection: Choose the appropriate engine based on audio language (e.g., Mandarin, Chinese-English mix, Cantonese, English) from the engine cheat sheet; for example, use 16k_zh-PY (sentence recognition) or 16k_zh_en (flash recognition) for mixed Chinese-English audio.
  4. Routing Decision: Combine audio size and URL/local path to select the recognition method; public URLs default to file_recognize.py rec, avoiding unnecessary local processing.

Applicability and Caveats

  • No Substitution: If scripts fail, return errors; do not guess transcriptions using models to ensure data reliability.
  • Credential Security: In group chats, avoid sending SecretId, SecretKey directly; in private chats, warn about leakage risks as credentials may pass through LLM.
  • Environment Dependencies: FFmpeg/ffprobe is required for audio processing; scripts support auto-installation via ensure_ffmpeg.py, but seek user help if it fails.
  • Single Task Focus: Avoid modifying system configs (like ~/.bashrc) for one-time recognition; use current command injection instead.
  • Flash Recognition Limits: If self-check passes for sentence and file recognition but fails for flash, suggest "common with international site accounts or domestic accounts accessed overseas" as a possible reason, not an absolute conclusion.
  • Minimal Interruptions: Unless users must provide credentials or explicitly request manual configuration, avoid unnecessary confirmations to keep the workflow smooth.

Use Cases

  • Transcribe meeting recordings with Chinese-English discussions into text documents for further organization and distribution.
  • Process user-uploaded Cantonese voice feedback to extract content for improving product localization strategies.
  • Batch convert podcast audio files using asynchronous recognition to generate text for content analysis.
  • Process speech in short videos with Chinese-English mix to generate subtitles for accessibility.

Best For

  • Project managers who need to process multiple meeting recordings weekly and organize them into text minutes.
  • Localization engineers responsible for product localization, analyzing dialect user feedback to improve localization strategies.
  • Independent podcast creators who want to convert audio to text for SEO optimization.
  • Social media managers running short video accounts, needing to add subtitles to Chinese-English mixed videos.