NVIDIA vs Speechmatics

Programming toolkit and libraries for running general-purpose computation on NVIDIA GPUs.

Suite of AI technologies (speech, intelligence, animation models and plugins) for building conversational, actionable in-game characters, mostly MIT-licensed with some components under NVIDIA's open model license.

Supported software suite for building and deploying AI workloads on NVIDIA hardware.

Free development-environment manager for creating, customizing and collaborating on AI applications across GPU systems, with enterprise support available through an NVIDIA AI Enterprise license.

Development platform for AI-driven biology and drug discovery: open models, libraries, datasets and NIM microservices for the full AI lifecycle in biopharma.

End-to-end autonomous-vehicle platform spanning training infrastructure, simulation and safety-certified in-vehicle compute for production autonomy from L2++ to L4.

Open-source model-deployment server for TensorRT, PyTorch, ONNX, OpenVINO and other frameworks across GPUs and CPUs, with production support and stable APIs through NVIDIA AI Enterprise.

Open robotics development platform combining simulation, CUDA-accelerated perception libraries, ROS 2 integration and humanoid tooling for building AI-powered robots.

Suite of libraries and microservices covering the AI agent lifecycle - data curation, customisation, evaluation, guardrails and monitoring.

Platform of OpenUSD-based libraries and microservices for building simulation-ready digital twins and synthetic-data environments used to train and validate physical AI and robots.

Suite of open-source, GPU-accelerated data-science libraries (cuDF, cuML, cuGraph, cuxfilter), rebranding to "CUDA-X for Data Science."

Deployable speech-AI library (ASR/TTS) for production inference, free to prototype via build.nvidia.com and licensed for production deployment through NVIDIA AI Enterprise.

Ecosystem of compilers, runtimes and optimization tools for high-performance deep-learning inference; TensorRT-LLM and Model Optimizer are free on GitHub, with commercial deployment options via NVIDIA AI Enterprise.

NVIDIA's AI cloud of GPU clusters and managed AI services, which NVIDIA's own page now presents as its internal environment for building its models, and which is still sold on Microsoft's Azure marketplace.

Prebuilt inference microservices that package foundation models with optimised serving engines and standard APIs, deployable as hosted endpoints or self-hosted containers.

Data-centre GPU architecture used for training and serving large models, sold in servers and rack systems.

Data-processing-unit (DPU) platform combining compute with accelerated networking, storage and security for AI-factory infrastructure.

Free application that enhances livestreams and video calls with AI noise removal, background replacement and other audio/video effects; requires a GeForce RTX GPU.

Desktop AI supercomputer powered by the GB10 Grace Blackwell Superchip for local autonomous agents, running models up to 200B parameters for inference.

Deskside AI supercomputer powered by the GB300 Grace Blackwell Ultra Superchip, up to 20 petaFLOPS and 748GB coherent memory, supporting models up to 1T parameters.

Reference baseboard specification (HGX B200, B300, Vera Rubin NVL8) that system integrators and OEMs build into complete AI servers, listed in NVIDIA's Qualified System Catalog.

Edge-AI developer kits (IGX Thor, IGX Thor Mini, IGX Orin) for industrial and medical edge computing.

Family of embedded AI modules and developer kits for robotics and edge AI (Thor, AGX Orin, Orin NX/Nano, Xavier, TX2/Nano series).

Scalable data-centre server systems (e.g. OVX L40S with four or eight GPUs) for AI and graphics workloads, built by NVIDIA OVX partners.

Quantum-X800 InfiniBand switching for large-scale AI clusters, part of NVIDIA's InfiniBand networking line.

Professional GPU line (RTX PRO 2000 through RTX PRO 6000) for AI, graphics and simulation workloads across Blackwell, Ada Lovelace, Ampere and Turing architectures.

Ethernet networking platform (switches, ConnectX SuperNICs, BlueField DPUs, LinkX cables) purpose-built for generative-AI-scale data centres.

NVIDIA's data-centre platform succeeding Blackwell, pairing Rubin GPUs with Vera CPUs at rack scale for agentic AI and reasoning workloads.

Software that virtualizes GPUs across VMs in enterprise data centres and cloud deployments, licensed through NVIDIA's vGPU License and Support Portal.

Speechmatics' automatic speech recognition API transcribes audio into text in 55+ languages in either real-time streaming or batch mode, with speaker diarization, custom dictionary, translation and summarization options.

Generates a short summary of an audio file in the same API call that transcribes it, as paragraphs or bullets. It is a Speech Intelligence feature enabled by adding a config block to a batch Speech to Text job, not a product bought on its own.

Speechmatics' text-to-speech API generates streaming synthetic English speech from text with sub-150ms latency using four named voices (Sarah, Theo, Megan, Jack), aimed at real-time voice agent use.

Translates a transcript into other languages in the same API call that produces it, for files or live audio. It is a feature switched on inside a Speech to Text request rather than a separate product; the docs file it under Speech to Text and return the translations alongside the transcript.

Speechmatics' voice agent offering provides a real-time conversational speech API - including the Flow WebSocket endpoint that chains speech-to-text, an LLM, text-to-speech and function calling - plus a Python Voice SDK for turn detection and speaker management, and integrations with Vapi, LiveKit and Pipecat.

A locally-executing speech-to-text engine for Mac and Windows laptops that runs on about one CPU core plus the device's neural engine or GPU and roughly 800MB of memory, sending no audio over a network and claiming accuracy within 5% of the cloud API.