[KongchangAI]
Model

Marimba

A unified model for multi-task speech decision-making built on a frozen ASR backbone. Task-specific readout layers enable typed speech decisions such as vocabulary detection, end-of-turn determination, language identification, and speaker-change detection.

Timeline (last 90 days)

Nov 17

Marimba is open-sourced under the Apache-2.0 license.

Unverified50%
Nov 17

Marimba streams audio every 250 milliseconds and can run locally on CPU or GPU.

Unverified50%
Nov 17

Marimba gives typed answers (yes/no, single-choice, multi-choice, calibrated scores) for a recent 8-second audio window.

Unverified50%
Nov 17

Marimba's installation command is pip install "marimba-audio[torch]".

Unverified50%
Nov 17

On the turn-end detection task, Marimba reaches an AUROC of 0.935, while the comparison models (Qwen3-Omni-30B-A3B, Qwen2-Audio-7B, FireRedAudio 10.6B) range between 0.54 and 0.79.

Unverified50%
Nov 17

Marimba performs poorly on the Speech Commands keyword detection task.

Unverified50%
Nov 17

Marimba wins in binary term detection on Common Voice and FLEURS.

Unverified50%
Nov 17

Marimba's turn-end detection results belong to an in-distribution task, because the training data included Smart Turn data.

Unverified50%
Nov 17

Marimba can be integrated into Pipecat, LiveKit, and Vapi for use as a turn detector.

Unverified50%
Nov 17

Marimba's current capabilities are limited to the question types it was trained on; for tasks requiring complex audio reasoning, large models remain the more suitable choice.

Unverified50%

11 more timeline events

All Facts (20)

Unverified

Marimba is open-sourced under the Apache-2.0 license.

50%
Unverified

Marimba streams audio every 250 milliseconds and can run locally on CPU or GPU.

50%
Unverified

Marimba gives typed answers (yes/no, single-choice, multi-choice, calibrated scores) for a recent 8-second audio window.

50%
Unverified

Marimba's installation command is pip install "marimba-audio[torch]".

50%
Unverified

On the turn-end detection task, Marimba reaches an AUROC of 0.935, while the comparison models (Qwen3-Omni-30B-A3B, Qwen2-Audio-7B, FireRedAudio 10.6B) range between 0.54 and 0.79.

50%
Unverified

Marimba performs poorly on the Speech Commands keyword detection task.

50%
Unverified

Marimba wins in binary term detection on Common Voice and FLEURS.

50%
Unverified

Marimba's turn-end detection results belong to an in-distribution task, because the training data included Smart Turn data.

50%
Unverified

Marimba can be integrated into Pipecat, LiveKit, and Vapi for use as a turn detector.

50%
Unverified

Marimba's current capabilities are limited to the question types it was trained on; for tasks requiring complex audio reasoning, large models remain the more suitable choice.

50%
Unverified

Marimba proposes using a frozen Qwen3-ASR 0.6B feature extractor as the backbone, paired with lightweight task-specific readout layers, to handle multiple speech decision tasks in a unified way.

50%
Unverified

Marimba's turn-end detection readout reaches AUROC 0.930, while the dedicated model Smart Turn v3.2 achieves 0.982 on its own test set.

50%
Unverified

In Marimba's readout logic, vocabulary-type questions are scored zero-shot directly on the ASR-transcribed text, with no additional classifier training required.

50%
Unverified

Marimba's language identification reuses the language scoring built into the ASR model, adding no extra module.

50%
Unverified

Marimba's turn-end judgment is read from layer-18 features of the backbone network, and the speaker-related readout is extracted from layer-12 features.

50%
Unverified

Marimba achieves decoder AUROC of 0.80–0.92 on a private conversation test set and 0.92–0.97 on public read speech.

50%
Unverified

Marimba achieves 0.93 accuracy on the four-class Speech Commands task.

50%
Unverified

Marimba's binary vocabulary judgment has a maximum ECE of 0.08.

50%
Unverified

Marimba's personal information detection achieves an entry-level F1 of 0.904 on synthetic Brazilian Portuguese speech and 1.000 on human reference transcripts.

50%
Unverified

Marimba's speaker-change detection achieves AUROC of 0.966–0.980 on AMI and VoxConverse with the layer-12 readout, outperforming the ECAPA baseline.

50%

Source Articles