DeepSeek API Adds Search, OCR and Speech: 120-Second Audio Cap, Vendor Bound by Chinese Laws
Reported by letsdatascience and Tech Times2 sources ↓
DeepSeek's Nov. 13 API update adds search, URL fetching, file parsing, OCR and speech tools. Speech recognition is limited to 120 seconds per request, and the guides name no model behind OCR or voice services.
Key points
- Chat features run on deepseek-flash across DeepSeek's Messages, Responses and Chat Completions APIs
- File parsing can miss charts and diagrams; rich-layout mode crops detected figures, which add input tokens
- Speech recognition accepts WAV or Ogg Opus files up to 8 MiB and does not support streaming
- The OCR guide lists a 100 MiB and 1,000-page limit; processing must finish within 180 seconds
Sources
Details
The changelog also lists a safety-check setting, standalone search and fetch tools, and an open-source demo chat client that can import a user's DeepSeek chat history. Tech Times reports that DeepSeek handles tool calls and context assembly on its servers; developers receive the resulting content blocks and pass them into the next request. Each successful search or fetch call is billed.
For file input, developers upload files and reference them by ID in chat messages. Text recognition extracts text but omits charts, photos and diagrams; rich-layout recognition can crop figures, though DeepSeek warns that not all figures will be detected. The OCR service returns page-by-page text with position coordinates, and its guide lists support for PDFs, Word and PowerPoint files and common image formats.
The speech guide lists 21 languages for automatic recognition and says stereo input uses only the left channel. Speech synthesis offers the stella and tide voices in 10 languages, with requests capped at 10,000 billed characters; the guides do not name underlying models for OCR or speech services. Tech Times also reports that calls go to a Hangzhou-based company subject to Chinese state-security laws.
Written from coverage as of