Compare realtime APIs

Pick up to four APIs from any category and see them side by side.

API 1API 2API 3API 4

	WhisperLive (Collabora) Collabora (open source)	NVIDIA Nemotron ASR Streaming / Parakeet (Riva, NIM, NeMo) NVIDIA	Kyutai STT (Delayed Streams Modeling) Kyutai
Category	Open models	Open models	Open models
Status	GA	GA	GA
Est. per minute	$0	$0	$0
How that was worked out	Self-host compute only.	Self-host compute only; cost depends on GPU price and streams per GPU.	Self-host compute only.
Pricing model	free	free	free
Free tier	Open source.	Open weights; hosted API trial on build.nvidia.com.	Open weights.
Connects by	WebSocket, REST (optional)	gRPC (Riva/NIM), WebSocket (Riva realtime client), Python (NeMo)	WebSocket (moshi-server), Python (PyTorch, MLX)
Audio in	Client captures mic or file and streams 16 kHz float audio.	16 kHz mono PCM typical for Riva streaming (not re-verified per model).	Handled by the provided scripts (mic or file); server streams PCM over WebSocket.
Audio out	n/a	n/a	n/a
Languages	Whisper's ~99 languages (model-dependent).	English (Nemotron 3) or 40 language-locales (Nemotron 3.5).	English and French (1B) or English (2.6B).
Latency (vendor claim)	Depends on model size and GPU; Whisper is not natively streaming, so partials are re-decoded windows.	Chunk sizes down to 80 ms per model card; NVIDIA FAQ reports 0.067 s ASR latency at 64 parallel streams for Parakeet CTC 1.1B on 3xH100 (vendor benchmark).	Model-defined delay: 0.5 s (1B) or 2.5 s (2.6B) per README.
Key limits	Server-side max_clients and max_connection_time settings cap concurrency and session length.	GPU required for the multilingual model. Third-party review notes occasional misplaced punctuation in streaming and inconsistent auto language detection; specify the language when known.	Fixed algorithmic delay per model; 2.6B is too slow-to-final for snappy voice agents.
High-severity warnings	None	None	None
Compliance	Self-hosted.	Self-hosted: data stays in your environment.	Self-hosted.
Self-hostable	Yes	Yes	Yes
Last checked	2026-10-10	2026-10-10	2026-10-10
Key numbers and features
Type	Speech-to-text	Speech-to-text	Speech-to-text
Params B	-	0.6	1
VRAM GB	-	-	-
CPU ok	Yes	No	-
Licence	MIT	OpenMDW-1.1	CC-BY-4.0
Commercial	Yes	Yes	Yes
Full duplex	-	-	-
Streaming	Yes	Yes	Yes
Latency ms	-	-	500
Languages	99	40	2
High warnings	0	0	0