Spaces:
Running on CPU Upgrade
New reliable tasks for narrow-domain ASR benchmarking
How can be added new tasks to this leaderboard?
I am Pablo Segovia Martínez, a member of the Machine Learning and Language Processing (MLLP) research group at the Universitat Politècnica de València (UPV). We recently published an English speech corpus of high-energy particle physics talks (235 hours), where 30 hours of these talks were manually transcribed to build two reliable tasks for narrow-domain ASR benchmarking. This corpus (called LHCP-ASR) has been published in a long paper accepted at Interspeech 2025, it can be read here.
These talks were extracted from 2020-2022 Large Hadron Collider Physics (LHCP) conferences and the 30 hours mentioned above were splitted into two sub-tasks, LHCP-2020 and LHCP-2022, with their respective development and test partitions:
- LHCP-2020:
- Dev: 5.6 h
- Test: 6.2 h
- LHCP-2022:
- Dev: 4.8 h
- Test: 13.4 h
This corpus is available on Hugging Face via this link https://huggingface.co/datasets/mllp/LHCP-ASR and it has been released in two formats: segments of the talks (less than 30s each) and longform, which provides the full talk.
Why are these tasks interesting? Each talk is a presentation followed by a question-and-answer session with multiple speakers, mostly non-native, under heterogeneous acoustic conditions and high audio compression. These conditions make LHCP-ASR a challenging task for ASR.
