| CueLight |
| Copyright 2026 sshah03 |
|
|
| Licensed under the Apache License, Version 2.0 (see LICENSE). |
|
|
| CueLight's models were trained on, and build on, the third-party work below. Each is used under its own license. |
| Where weights or data were changed (fine-tuned, re-segmented, relabeled or used for training), that's noted. |
|
|
| Model weights |
|
|
| - CueLight-Base is fine-tuned from the encoder of OpenAI's Whisper small.en |
| (https: |
| (made causal) and fine-tuned. Whisper's code is Copyright (c) 2022 OpenAI, under the MIT License |
| (LICENSES/whisper-MIT.txt). |
| - Whisper base.en and medium.en (OpenAI, Apache-2.0 weights and MIT code) were used as teachers during training. |
| - Smart Turn v3.2 (https: |
| License (LICENSES/smart-turn-BSD-2-Clause.txt). No Smart Turn code is included here. |
|
|
| Training data (Creative Commons Attribution, used for training after re-segmenting and relabeling) |
|
|
| - Smart Turn v3.2 training data, pipecat-ai/smart-turn-data-v3.2-train, CC BY 4.0, by Pipecat / Daily, with audio |
| contributed by Liva AI, Midcentury and MundoAI. |
| https: |
| - The People's Speech (clean and dirty subsets), MLCommons, CC BY. Galvez et al., "The People's Speech: A Large-Scale |
| Diverse English Speech Recognition Dataset for Commercial Usage", 2021. |
| https: |
| - Harper Valley Bank, Gridspace and Stanford University, CC BY 4.0. |
| https: |
| - Open Yap 1K sample, The Agentic Data Company, CC BY 4.0. |
| https: |
| - eot-bench data, LiveKit, CC BY 4.0, used to choose between models and for evaluation. |
| https: |
|
|
| Conditions passed on from the data: please don't use CueLight, or audio processed with it, to identify speakers or to |
| clone voices. |
|
|
| License texts: https: |
|
|
