VoiVision AI

Tech & Ecosystem

Four hardware lines · one engine · full coverage

AI Speech Engine

Built on NEU NLP Lab's in-house research, we re-architected the ASR pipeline from the traditional AM+LM+Decoder cascade into an end-to-end (E2E) neural model. Using Conformer + CTC/Attention hybrid modeling, combined with INT8/INT4 quantization, pruning, knowledge distillation, structural re-parameterization, and operator fusion, the engine achieves real-time ASR on CPU alone — no GPU required. Lightweight language modeling with adaptive context compression reduces LM compute by over 70%, while recognition accuracy improves 10–20% over traditional approaches.

  • ▸E2E model: Conformer + CTC/Attention hybrid, eliminating acoustic/language modeling error accumulation
  • ▸INT8/INT4 quantization & pruning: drastically reduced params for real-time CPU inference
  • ▸Knowledge distillation & structural re-parameterization: compact model with high generalization
  • ▸Operator fusion & cache optimization: improved CPU matrix efficiency
  • ▸Lightweight LM: internal LM + adaptive context compression, 70% compute reduction
  • ▸Context-enhanced training & constrained self-attention: high accuracy in low-resource conditions

© VoiVision Speech Transcription Meeting Software V1.0

Accelerator Support

  • •NVIDIA: T4 / L4 / A10 / A100
  • •Ascend: 310P / 910B
  • •Cambricon: MLU370 / MLU590
  • •Hygon DCU (CPU-only also works)

Full OA Integration

  • •WeCom
  • •DingTalk
  • •Feishu
  • •Landray
  • •Seeyon

Deployment

  • •Docker / K8s one-click
  • •Bare-metal
  • •Gov cloud / private cloud / SIEM
  • •RESTful API / WebSocket
Explore EngineTech Inquiry

22 Chinese Dialects · Auto-Detected

Covers 22 Chinese dialects: 15 Mandarin branches, plus Wu, Min, Gan and Yue (Cantonese / Hong-Kong Cantonese). The system auto-detects language and dialect, handles mixed Chinese-English speech with no manual switching, and covers meetings that mix Mandarin with dialect.

Dialect Coverage

  • ▸15 Mandarin branches (Northeastern / Beijing / Jilu / Zhongyuan / Lan-Yin / Jiang-Huai, etc.)
  • ▸Wu (Jiangsu-Zhejiang-Shanghai)
  • ▸Min (Southern Min / Eastern Min / Teochew)
  • ▸Gan (Jiangxi)
  • ▸Yue (Cantonese / Hong-Kong Cantonese)

Three Paths to Higher Accuracy · Stackable

From a general ASR to industry-grade precision, accuracy can be lifted step by step along three paths, combined freely by rollout pace:

  • ① Screen OCR (zero-config): boost recognition with on-screen content — term accuracy 72% → 94%+
  • ② Hotword injection (minutes): add domain terms without retraining ASR; hotwords injected into the decoder in < 500ms
  • ③ Vertical model (weeks): fine-tune on customer corpora to better fit business terms and phrasing

LLM Post-Correction · +5–10% Overall

Transcripts are processed a second time by an LLM: error correction, smoothing, auto-punctuation and number normalization, plus RAG augmentation from the enterprise knowledge base for proper nouns and business terms — lifting overall accuracy a further 5–10%.

  • ✓Error correction: fix homophones and transcription errors
  • ✓Smoothing: merge fragments and tidy spoken phrasing
  • ✓Auto-punctuation: restore sentence breaks and paragraphs
  • ✓Number normalization: unify amounts / dates / IDs
  • ✓RAG augmentation: correct proper nouns against the enterprise KB

ASR × OCR Multimodal Fusion

Screen-content awareness × dynamic ASR enhancement

VoiVision links ASR with on-device OCR, using the visual content of the screen (PPT / docs / whiteboard) to boost speech recognition. The BJ66+WCB06 captures the screen via wired + wireless channels; an on-device large-model OCR reads the text frame by frame; the domain knowledge base auto-matches and injects domain terms into the ASR decoder — lifting accuracy in terminology-dense scenes like lectures, consults, and roadshows.

Four-step closed loop

  1. ▸① Capture: BJ66+WCB06 grabs the screen byte stream (HDMI-in + wireless)
  2. ▸② On-device OCR: frame-by-frame parsing of PPT / docs / whiteboard text
  3. ▸③ KB match: 200+ discipline hotword dictionaries auto-align
  4. ▸④ Hotword injection: domain terms dynamically injected into ASR decoder

Key advantages

  • ✓Fully local private deploy — data never leaves room
  • ✓Hot-swappable domain KBs (education / healthcare / finance / law)
  • ✓New domains need no ASR retraining — hotwords take effect instantly
  • ✓Hotword injection < 500ms, frame-level OCR < 500ms
Generic ASR term accuracy baseline ~72% → ASR×OCR fusion lifts to 94%+ (NEU NLP Lab lecture-hour test)

OA Integration

Seamless with mainstream office platforms · minutes into your workflow

VoiVision deeply integrates with leading office & collaboration platforms — automatic minutes sync, smart todo extraction, one-terminal multi-platform management — embedding meeting intelligence directly into existing enterprise workflows.

WeCom

  • •Calendar sync · contacts
  • •Approval flow · app push
  • •Auto minutes sync · todo

DingTalk

  • •Calendar sync · meeting mgmt
  • •DING notify · approval
  • •Todo · multi-dimension sheets

Feishu

  • •Calendar sync · docs
  • •Msg notify · meeting
  • •Approval · enterprise wiki

Landray / Seeyon

  • •Eco integration · SSO
  • •Workflow engine link
  • •Custom self-built OA

Book a Personalized Demo

Tell us your meeting scenario and compliance needs — get a tailored plan.

Book now
Live Chat