.wave documentation
.wave provides APIs for real-time speech processing, including full-duplex voice conversations, streaming transcription in OpenAI format, and a Deepgram-compatible API.

Wave sells dedicated inference for real-time AI with guaranteed latency and streaming speech and conversation models.
dotwave.aiCrawled Nov 11, 2026
.wave provides APIs for real-time speech processing, including full-duplex voice conversations, streaming transcription in OpenAI format, and a Deepgram-compatible API.
Reference for the .wave APIs, grouped by API: the Live API, the Realtime API, the Deepgram-compatible API, and the models, sessions and usage routes they share.
.wave offers reserved inference capacity with guaranteed latency, an SLA, and a fixed upfront price, using dedicated lanes instead of whole GPUs.
Nemotron 3.5 ASR Streaming 0.6B on the .wave Live API, in public beta: NVIDIA’s streaming speech recognition in 32 language locales, at $0.00045 per minute.
Nemotron VoiceChat 11B on the .wave Live API, available on request: NVIDIA’s full-duplex English speech-to-speech model, with live transcripts of both speakers, at $0.005 per minute.
Hosted streaming models for live speech and conversation: streaming speech-to-text and full-duplex voice on .wave, with no GPU reservation required.
The company compiles continuous inference workloads into a persistent GPU program to handle recurring tasks with deadlines, achieving 56 times more real-time sessions on an H100.
What full-duplex voice reveals about continuous inference: persistent sessions, cadenced batching, timing constraints, and useful session density. A .wave article.
Stream microphone audio over a plain WebSocket on the .wave Live API and play the model’s speech, in Python.
How .wave serves 4,800 concurrent Nemotron 3.5 ASR streams on one H100 at the 80 ms chunk, twenty times the model card's published concurrency, at $0.00045 per minute, against $0.0045 for the same weights on Together AI. Results, business impact and measurement conditions.
How .wave serves 56 concurrent Nemotron VoiceChat 11B conversations on one H100, with 98% lower GPU cost per conversation at equal GPU price and utilization. Results, business impact and measurement conditions.
Prices per model, free credits, session limits and project limits on .wave.
Dedicated inference with guaranteed latency. Streaming models for speech and conversation.
The terms that govern self-serve access to the .wave real-time inference API.