Running Meeting Transcription On-Prem on Ascend & Cambricon
From why localization matters to a supported-accelerator list and containerized deployment, a practical playbook for running VoiVision's transcription engine on domestic hardware in private environments.
Why Localization Matters
In government, finance, and defense, systems must be not only "data on-premises" but also "technologically self-reliant." That means shifting the underlying compute from imported GPUs to Ascend, Cambricon, and Hygon domestic chips, and the OS to Kylin, UnionTech, and other domestic systems.
Core point: Private deployment answers "where the data is"; localization answers "who controls the technology." Together they make truly autonomous meeting intelligence.
Supported Accelerators
Built on NEU NLP (Northeastern University) research, the VoiVision engine natively supports mainstream accelerators:
| Type | Representatives | Use |
|---|---|---|
| NVIDIA | T4 / L4 / A10 / A100 | General, high performance |
| Ascend | 310P / 910B | Localized, domestic servers |
| Cambricon | MLU370 / MLU590 | Localized, domestic servers |
| Hygon DCU | DCU series | Domestic x86 path |
| CPU only | x86 / ARM | Reuse existing, low concurrency |
Containerized One-Command Deployment
Shipped as a container image with Docker and Kubernetes support:
- Runs on bare metal, government cloud, and localized environments;
- Integrates with OA and meeting systems via standard RESTful / WebSocket APIs;
- Scales elastically by concurrency.
If you already have servers, deploy the software-only engine for private transcription on domestic accelerators without buying hardware.
Accuracy and Performance
Localization is not a downgrade. The algorithm relies on accelerators only for inference; after quantization and operator adaptation:
- Chinese transcription accuracy stays at the same level (VV01 offline host >95%);
- Latency under 3 seconds;
- Throughput scales linearly with accelerator specs.
Sizing Advice
- New localized cluster: VV50 / VV10 + Ascend or Cambricon accelerator;
- Reuse existing x86: engine on CPU or Hygon DCU;
- Mixed environment: unify NVIDIA and domestic accelerators, schedule by load.
See the private deployment guide and MLPS Level 3 compliance.
FAQ
Q: Which domestic chips does transcription support?
A: The VoiVision engine natively supports Ascend (310P/910B), Cambricon (MLU370/MLU590), Hygon DCU, and the full NVIDIA lineup; it also runs on CPU alone, covering the entire localized and domestic stack.
Q: Can speech transcription run on CPU only?
A: Yes. The engine is inference-optimized for CPU, so it runs on localized or general servers without a discrete accelerator—ideal for low concurrency or reusing existing hardware.
Q: How is it deployed with Docker / K8s?
A: VoiVision ships as a container image with one-command Docker and Kubernetes deployment, running on bare metal, government cloud, and localized environments, and integrates via standard RESTful / WebSocket APIs.
Q: Does accuracy drop on domestic accelerators?
A: No. The model is built on NEU NLP (Northeastern University) research; accelerators only affect throughput and latency. After quantization and operator adaptation on Ascend/Cambricon, Chinese transcription accuracy stays at the same level.