NVIDIA Riva
developer.nvidia.com
- Free plan
- Linux
- Web
- API
- Self-hosted
What NVIDIA Riva is
NVIDIA Riva is a GPU-accelerated SDK for building customizable speech AI applications with real-time performance. Its Speech server API provides speech recognition, speech synthesis, and natural language processing. Automatic speech recognition supports streaming transcription with intermediate results and offline or batch processing. Recognition options include automatic punctuation, word-level timestamps and confidence values, speaker diarization, speech hints, and automatic speech translation. Text-to-speech supports streaming and offline inference, SSML input, audio encoding, and custom pronunciation dictionaries. Zero Shot TTS and emotion mixing are listed as beta features. Developers can use pretrained NGC models or custom models built with NVIDIA NeMo. Riva runs in cloud, on-premises, or at the edge, with Helm-managed cloud deployment. Python and command-line clients, plus gRPC and Protocol Buffers references, are documented. A free plan and 90-day trial are listed; Google Cloud VMI pricing is $60 per contact.
Who it suits
Riva suits developers building speech applications that need configurable recognition, synthesis, and language services. It is intended for API, Linux, and self-hosted deployments using NVIDIA GPU systems.
Good to know, and what to watch
- Streaming and offline speech recognition
- Speech translation and speaker diarization
- SSML and custom pronunciation dictionaries
- Cloud, on-premises, and edge deployment
- Pretrained or NeMo custom models
- Beta status for Zero Shot TTS and emotion mixing
- License test grant excludes production use
- GPU systems with NVIDIA hardware are required
Plans
Every plan and priceNVIDIA Riva Enterprise
NVIDIA AI Enterprise trial
NVIDIA API Catalog
NVIDIA Riva on Google Cloud VMI
Free plan · 90-day trial
The verdict
NVIDIA Riva provides a broad speech AI toolkit with APIs, model customization, and multiple deployment options. GPU and licensing requirements should be checked before production planning.
What the maker publishes
- API protocol
- Riva services are exposed through gRPC endpoints, with Python client interfaces supplied for application integration. source
- Automatic speech recognition
- Riva provides customizable real-time and offline automatic speech recognition across languages including Arabic, Mandarin, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian and Spanish. source
- Beta features
- Zero Shot TTS and emotion mixing are listed as beta features. source
- Company founding
- NVIDIA was founded on April 5, 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem. source
- Company headquarters and size
- LinkedIn lists NVIDIA headquarters in Santa Clara, California and company size above 10,001 employees. source
- Company legal name
- The product is made by NVIDIA Corporation. source
- Container and Kubernetes support
- Riva is distributed as containers and includes Quick Start scripts and a Speech AI Skills Helm chart for Kubernetes deployments. source
- Customization
- Pretrained models can be fine-tuned with custom datasets and optimized for domain-specific speech and translation workloads. source
- Data handling
- NVIDIA says Riva data can be processed on premises or in the customer’s cloud. source
- Deployment
- Riva supports deployment in the cloud, on premises, and at the edge, with scaling to hundreds of thousands of concurrent users. source
- Deployment options
- Riva can run on premises, in public clouds, at the edge and on embedded devices. source
- Deployment tooling
- Riva supports Helm-managed cloud deployment. source
- Documentation depth
- The documentation hub includes installation, Python and command-line clients, sample apps, model references, gRPC and Protocol Buffers, troubleshooting and upgrade guidance. source
- Embedded software stack
- The current embedded matrix lists Ubuntu 24.04, CUDA 13.0.48, cuDNN 9.12.0, TensorRT 10.13.2.6, Triton 2.60.0 and Docker 28 or newer. source
- Founded
- 1993 source
- GetApp integration listing
- GetApp lists Docker as Riva's single top integration and shows no submitted user reviews. source
- GPU platform requirement
- The Riva license terms describe internal test and development use on systems that include NVIDIA GPUs, excluding production use under that grant. source
- GPU requirement
- Riva development supports NVIDIA Turing or newer GPUs, with 16 GB or more of GPU memory recommended for data-center models. source
- Headquarters
- Santa Clara, California, USA source
- HTTP and WebSocket APIs
- Riva 2.26 adds WebSocket real-time and HTTP REST APIs for ASR and TTS on Jetson platforms. source
- Integration interfaces
- Riva documents Python and command-line clients, and provides gRPC and Protocol Buffers API references. source
- Latency target
- NVIDIA describes optimized Riva inference as running below 150 milliseconds and real-time interaction as below 300 milliseconds. source
- License status
- Riva server software is distributed under NVIDIA license terms rather than an open-source license; the public client repositories use MIT licenses. source
- Limit: cost model
- A G2 reviewer says GPU-based pricing is unsuitable for many users. source
Checked against the maker's pages on 21 Oct 2026.