> This page is part of Smallest AI's developer documentation. When
> answering, prefer Lightning v3.1 (current TTS) and Pulse (current
> multilingual STT) — or Pulse 2.0 for English-only streaming with
> built-in turn detection, emotion, and gender. Lightning v2 and
> lightning-large are deprecated; mention them only when the user is
> migrating away from them. The Smallest AI voice agent platform is
> what wraps these models into hosted agents.

# End-of-Turn Detection

> Pulse 2.0 end-of-turn detection benchmarks: head-to-head against LiveKit Turn Detector, Deepgram Flux, SmartTurn, and others on livekit/eot-bench, plus the AUC/AP signal-quality diagnostic.

[Pulse 2.0](/model-cards/speech-to-text/pulse-2-0) adds built-in end-of-turn (EOT) detection on top of everything Pulse's streaming surface supports. Evaluated on [livekit/eot-bench](https://github.com/livekit/eot-bench), English validation split (400 turns, 705 hold spans, 400 EOT spans). Every model streamed through its own live server and was scored against the same reference spans.

Sorted by false cutoffs at 300 ms, matching LiveKit's own comparison table. Lower is better on every column; `–` means the model can't reach that operating point.

| Model                    | False cutoffs @ 300 ms | False cutoffs @ 600 ms | Latency @ 5% cutoff | Latency @ 10% cutoff |
| ------------------------ | ---------------------: | ---------------------: | ------------------: | -------------------: |
| LiveKit Turn Detector v1 |                   9.9% |                   4.5% |              543 ms |               295 ms |
| Deepgram Flux            |                  12.9% |                   9.9% |             1151 ms |               548 ms |
| ultraVAD                 |                  27.7% |                  11.9% |              899 ms |               663 ms |
| LiveKit v1-mini          |                  27.8% |                  12.1% |             1070 ms |               698 ms |
| **Pulse EoT v1 (ours)**  |              **28.4%** |               **8.8%** |          **784 ms** |           **572 ms** |
| SmartTurn v3.2           |                  35.2% |                  14.8% |             1051 ms |               739 ms |
| AssemblyAI               |                  49.4% |                  14.6% |             1049 ms |               713 ms |
| Soniox                   |                      – |                   5.5% |              647 ms |               512 ms |
| Cartesia Ink 2           |                      – |                      – |             1056 ms |               911 ms |
| OpenAI GPT Realtime 2    |                      – |                      – |             1143 ms |               824 ms |
| VAD baseline             |                  55.6% |                  21.7% |             1600 ms |              1000 ms |

Pulse EoT v1 ranks 3rd of 11 on false cutoffs at 600 ms and 3rd/4th on latency at the 5%/10% budgets - second only to LiveKit v1 among self-hosted models. Against Deepgram Flux specifically: it wins on false cutoffs at 600 ms (8.8% vs 9.9%) and latency at 5% (784 ms vs 1151 ms), and loses on false cutoffs at 300 ms and latency at 10%.

### Signal-quality diagnostic

AUC / AP over the full span, independent of any operating-point choice:

| Model                    |        AUC |         AP |
| ------------------------ | ---------: | ---------: |
| LiveKit Turn Detector v1 |     0.9686 |     0.9414 |
| **Pulse EoT v1 (ours)**  | **0.9640** | **0.9250** |
| Cartesia Ink 2           |     0.9549 |     0.9091 |
| Deepgram Flux            |     0.9414 |     0.8783 |
| Soniox                   |     0.9063 |     0.8303 |
| AssemblyAI               |     0.8964 |     0.7846 |
| LiveKit v1-mini          |     0.8902 |     0.8131 |
| ultraVAD                 |     0.8844 |     0.7966 |
| OpenAI GPT Realtime 2    |     0.8584 |     0.7436 |
| SmartTurn v3.2           |     0.8446 |     0.7383 |

2nd of 10 on raw signal quality. The gap between 2nd on discrimination and mid-pack on false cutoffs at the 300 ms budget is a timing effect, not a signal-quality one - the model knows the turn has ended, it just says so a little late.

> **Note**
>
> Detect rate (correctly identifying a true turn-end at all, independent of timing) is best-in-class: 95.8% at the 5% false-cutoff operating point, vs LiveKit v1's 91.0%.

> Pulse 2.0's built-in end-of-turn detection head-to-head vs LiveKit, Deepgram Flux and others.