DeepSeek has officially opened its full application API suite, splitting web search, deep-thinking search, file upload, link reading, and for the first time speech recognition and synthesis into independently callable, separately billed endpoints. The company says developers can replicate the complete DeepSeek App 1:1 and has open-sourced a lightweight chat frontend demo. Unlike the October Search/Fetch APIs that required manual tool orchestration, the new endpoints let the model autonomously decide when and what to search.
ASR
自动语音识别
Technology that converts human speech into text, as demonstrated by models like Whisper.
Curated Intelligence
DeepSeek has exposed the full capability stack behind its consumer app through its API, letting developers enable web search, file and PDF parsing, link reading, safety filtering and voice with a few parameters, or call standalone OCR, ASR and TTS endpoints. An open-source frontend demo with conversation-history import support means developers can closely recreate the entire DeepSeek App.
Hugging Face's Open ASR Leaderboard has added private evaluation splits for Egyptian and Nigerian English accents, continuing its push to use held-out data for more trustworthy benchmarking. The update broadens comparable ASR coverage for under-represented English accents, but remains a routine dataset expansion without new models or methods.
Voice Arena has released a speech recognition benchmark aimed at real-world voice agent scenarios, with its first dataset, Monsoon, covering two-person conversations in five languages recorded on everyday phones. The benchmark stresses overlapping speech, interruptions, short backchannels and telephony compression, with a public leaderboard and test sets.
Cicada is a fully native, open-source macOS tool offering real-time bilingual subtitles, AI meeting minutes, and simultaneous interpretation. It integrates with major platforms like Zoom and Feishu via virtual microphones for cross-language communication, powered by Ling and Ming-3.0 models under the Apache-2.0 license.
Developer Joey showcased a fully offline home voice assistant prototype using DeepSeek V4.1 Flash for reasoning and Ming-3.0 models for ASR and TTS. Running on NVIDIA DGX Sparks and a Mac mini, the setup achieves low-latency interaction with zero cloud dependency. The code repository is not yet public, remaining in a personal demo stage.
NVIDIA showcased 'MonsterHouse', an interactive Halloween projection built on the Nemotron open model stack. The application integrates Nemotron ASR, Magpie TTS, and Audio2Face-3D to create a real-time voice agent with emotional feedback capabilities. This case serves primarily as a technical demonstration of multimodal component integration in specific scenarios.
Nvidia has released Nemotron 3.5 Transcribe, an open-source Automatic Speech Recognition (ASR) model supporting 40 languages. It integrates speaker attribution, translation, and audio Q&A capabilities, achieving top leaderboard results. The release also includes the Granary v2 data pipeline and fine-tuning recipes to facilitate developer customization.
Nvidia released the open-weight Nemotron 3.5 transcription model, supporting 40 languages and translation to/from English for 31. With 30B parameters, it runs on a single chip and handles long recordings with speaker diarization, aiming to lower barriers for developers building meeting and medical transcription apps.
Following the release of Anthropic's Opus 5.5, AI agents can now autonomously perform complex video editing tasks with significant efficiency gains. The model achieves 'code-based editing' by generating HTML/CSS/JS, drastically reducing reliance on external skills and specialized editors, signaling a shift towards automation in video production.