Basis releases 1,502 hours of multi-speaker audio, including a 15.3-hour Chinese subset

Basis released Conversations 1500 on September 23, offering 1,502 hours of multi-speaker conversations in 22 languages involving 2,645 speakers. The hours are measured along the conversation timeline, without counting each participant’s track separately. The data comes from participants talking with strangers who speak the same language through an app, supporting research into spoken interaction, turn-taking, and overlapping speech. [Release announcement](https://huggingface.co/blog/basis-ai/conversations-1500)

The dataset preserves a separate microphone track for each participant, with tracks in the same segment sharing the same start time and duration. This makes it easier to analyze who responds to or interrupts whom, and when. Calls accommodate two to four people at a time, but participants can change during a call, so a complete conversation may include more than four speakers in total. Audio samples are available on the official site, allowing researchers to examine the acoustic and interaction characteristics of natural conversation. [Project page](https://withbasis.co/open/conversations-1500)

The tracks use a lossless 48 kHz, 16-bit mono format, preserving the volume levels and noise present during recording, without uniform noise reduction or loudness normalization. During preprocessing, researchers should preserve the original time alignment to avoid compromising the detection of overlapping speech. [Audio specifications](https://huggingface.co/blog/basis-ai/conversations-1500)

Annotations are concentrated in a subset of approximately 100 hours. The team first identified candidate events using waveforms, audio models, and transcript analysis, then had humans review them. The events include backchannels, laughter, pauses, interruptions, and self-corrections. The announcement lists 23,751 human-reviewed moments. Except for laughter and backchannels, machine-generated labels for other event types are not guaranteed to have all been verified by humans, so evaluations should distinguish between machine-generated and human-verified labels. [Annotation process](https://huggingface.co/blog/basis-ai/conversations-1500)

Users working with Chinese should pay particular attention to the distribution: the dataset card lists only 15.3 hours, 17 conversations, and 35 speakers for the zh-TW subset. The human annotation hours field is blank, so the overall annotation coverage cannot be assumed to apply to Chinese. The official documentation also explicitly warns that automated transcripts may contain errors and are primarily intended for searching and filtering segments. It advises against using them as supervised training targets without first reviewing and correcting them. [Language and transcript details](https://huggingface.co/datasets/basis-ai/basis-conversations-1500)

Access still requires an application and manual review. According to the official documentation, the data is free for commercial and research use, subject to the dataset license and conditions that prohibit identifying, contacting, or impersonating specific speakers. For experiment design, we recommend splitting training and test sets by stable speaker IDs and creating a separate, manually annotated Chinese evaluation subset to reduce the effects of speaker overlap and annotation differences on results. These usage recommendations are based on the dataset’s structure; they do not represent demonstrated improvements in model performance. [Access conditions and data structure](https://huggingface.co/datasets/basis-ai/basis-conversations-1500)