Preprint
CS-Dialogue: A 104-Hour Dataset of Spontaneous Mandarin-English Code-Switching Dialogues for Speech Recognition
Jiaming Zhou et al., arXiv, 2025
Read the originalRead on October 17, 2026
- What it is
- A 2025 dataset paper that releases 104 hours of spontaneous Mandarin-English dialogue. It compares several pre-trained speech models on its test set, including Whisper Large-V2.
- How much weight it carries
- A preprint introducing a dataset, with the comparison run by its own authors on their own test set. Results on spontaneous dialogue are harder than on read speech, so the error rates describe this dataset, not a general benchmark.
Facts taken from this source
- On the CS-Dialogue Mandarin-English test set, Whisper Large-V2 had a 31.11 percent word error rate and a 15.29 percent mixed error rate, the highest error rates among the pre-trained models the authors tested.
“exhibits the highest error rates among the pre-trained models, particularly in WER (31.11%) and MER (15.29%)” (Results, Table 8 and surrounding text)