Baseten Diarized Transcription

We offer a multitude of high-performance, high-quality options for both pre-recorded and real-time diarized transcription. Each option is available as a single-tenant, dedicated deployment.

Diarized transcription answers who said what, taking audio input and returning a time-ordered transcript, in which each turn is attributed with a speaker. This basically enhances transcription with speaker metadata for meeting summaries, voice agents, or as a reviewable log.

The correctness of a diarized transcription is measured by cpWER, which incorporates the accuracy of both transcription and attribution.

Map of Baseten Diarized Transcription Offerings

Below is how our offerings compare among each other and against common closed-source APIs.

For recordings, accuracy and speed do not trade off the way you would expect: the most accurate option is also the fastest by a wide margin.

✕

Batch diarized transcription — accuracy against throughput

Realtime, the spread is wide as well. Our combinations land at roughly half the word error of the closed real-time APIs.

✕

Real-time diarized transcription — time to speaker label against accuracy

Get Started with Nemotron-3 Diarization + Qwen3-ASR

Our first recommendation combines NVIDIA's Nemotron-3 Diarization with a Qwen3-ASR-1.7B fine-tuned by Baseten. is the most accurate option we measured in batch and the fastest, transcribing at 187x real time, it handles more than thirty languages. This combination can also handle realtime audio over WebSocket from a second deployment, returning a speaker label 3.8 seconds after a speaker stops.

Try batch deployment · Deploy

For prerecorded files, a single six-minute meeting comes back in under four seconds, and throughput climbs steeply with concurrency before flattening at roughly 712x real time. Past that point additional concurrency buys queue depth rather than throughput, so latency grows while the work done per GPU stays flat. We recommend starting with RTX 6000 at concurrency 16.

✕

Batch performance — throughput and latency against concurrent requests

Over WebSocket, the finished transcript for a turn lands in about half a second after it is spoken. We are able to maintain this performance up to 56 streams per GPU. For a balance of latency and headroom across both, we recommend starting with RTX 6000 at concurrency 32 and adjusting from there. See the WebSocket concurrency guide for how to tune it.

Try real-time deployment · Deploy

✕

Realtime performance — time to final segment against concurrent streams per GPU

Both deployments are one click, and the request contracts are documented end to end: start with the batch and real-time quickstart, the streaming transcription API reference, and the speaker diarization overview. A runnable client lives in the Qwen3-ASR + Nemotron diarization cookbook, and the Parakeet combination has its own. For how we made these models fast, read how Baseten makes pyannote's diarization models 96x faster.

Alternatives on Baseten

Nemotron-3 + Parakeet is English-only by design and currently the most accurate option for live English audio, at a slightly later label than the Qwen combination. Deploy it for real-time, or for batch from the Nemotron 3 Diarization listing.

MOSS Transcribe-Diarize does both transcription and diarization with a single 0.9B model behind an OpenAI-compatible endpoint — one checkpoint to operate, at some cost in accuracy. See MOSS Transcribe Diarize.

pyannote community-1 + Whisper returns a provisional speaker label in about 1.3 seconds and revises it as the conversation disambiguates, for live captioning and speaker-change UI where a fast tag beats a correct one. It ships as a chain rather than a listing — talk to an engineer to set it up.

Contact us

Need this tuned to your audio, your languages, or your concurrency profile? Talk to an engineer.