Speech Engine (Software Only)
Conformer E2E · real-time on CPU · fully offline
Built on NEU NLP Lab's in-house engine with Conformer + CTC/Attention E2E architecture. INT8/INT4 quantization, pruning, knowledge distillation, and operator fusion enable real-time ASR on CPU alone. Supports 30 languages + 22 dialects, LLM translation/summary across 100 languages, machine translation at ≥800 chars/s, and 10:1 file transcription. Native support for NVIDIA (T4/L4/A10/A100), Ascend (310P/910B), Cambricon (MLU370/590), Hygon DCU; Docker/K8s one-click deploy.
Highlights
- ✓Conformer + CTC/Attention E2E architecture
- ✓Real-time on CPU · no GPU needed
- ✓INT8/INT4 quantization + pruning + distillation
- ✓LM compute −70% · accuracy +10–20%
- ✓30 languages + 22 dialects · LLM translation/summary for 100 languages
- ✓MT ≥ 800 chars/s · 10:1 file transcription
- ✓ASR×OCR multimodal fusion · term accuracy 94%+
- ✓NVIDIA / Ascend / Cambricon / Hygon · Docker/K8s · RESTful/WebSocket
Software Modules
Machine Translation
Text translation across 100 languages
- •Translation speed ≥ 800 chars/s
- •Multilingual translation incl. Chinese/English
Speech Recognition
Streaming, one-shot, and file transcription modes
- •Live streaming recognition with real-time text
- •One-shot utterance recognition
- •Audio file recognition with file transcription
LLM Minutes
LLM-powered structured meeting minutes
- •Chinese minutes: text and graphic formats
- •Automatic to-do extraction
- •Mind-map visualization
User-side API Integration
Bidirectional data flow with WeCom / OA
- •API for private minutes platform with WeCom and third-party OA systems
- •Display, query and manage full minutes inside WeCom / OA terminals
Book a Personalized Demo
Tell us your meeting scenario and compliance needs — get a tailored plan.