pilotgpt-test-v2
收藏资源简介:
pilotgpt-test-v2 是一个语音数据集,专为 Whisper 模型相关任务设计,使用 Trelis Studio 工具准备。该数据集包含 2 个源文件,共计 42 个验证样本,总时长为 4.6 分钟。数据经过 Silero VAD 处理,去除了静音部分,仅保留语音区域以确保与 faster-whisper 推理行为一致。数据集包含 8 个字段:音频片段(16kHz)、纯文本转录、带 Whisper 时间戳标记的转录、片段在原音频中的起止时间、语音持续时间(不含静音)、词级时间戳以及源文件名。特别适用于 Whisper 时间戳训练任务,建议采用双桶策略:50%样本使用纯文本转录,50%样本使用带时间戳标记的转录。
pilotgpt-test-v2 is a speech dataset tailored for tasks related to the Whisper model, prepared using the Trelis Studio tool. This dataset comprises 2 source files, totaling 42 validation samples with an overall duration of 4.6 minutes. The data was processed via Silero VAD to eliminate silent segments, retaining only speech regions to align with the inference behavior of faster-whisper. The dataset includes 8 fields: audio clips (16kHz), plain text transcriptions, transcriptions with Whisper timestamps, start and end times of segments within the original audio, speech duration (excluding silence), word-level timestamps, and source file names. It is particularly suited for Whisper timestamp training tasks, and the two-bucket strategy is recommended: 50% of the samples use plain text transcriptions, while the remaining 50% use transcriptions with timestamps.




