ghananlpcommunity/voxcpm-ghana-latents
收藏资源简介:
VoxCPM Ghana — Precomputed AudioVAE Latents 是一个用于文本转语音任务的数据集,专门用于微调模型 ghananlpcommunity/voxcpm-ghana。它包含预计算的 VoxCPM-0.5B AudioVAE 潜在表示(16 kHz),覆盖 42 种加纳语言(包括 Twi 的两种变体:twi-asante 和 twi-akuapem)以及过滤的加纳英语。数据集中每个剪辑都带有语言标签的转录文本,格式为 <|lang:CODE|>,其中 CODE 是 ISO-639-3 语言代码。数据集总计 1,756,157 个剪辑,约 3,400 小时音频,采样率为 16 kHz。AudioVAE 潜在表示的维度为 64,跳跃长度为 640(约每秒 25 帧)。数据以 parquet 分片格式存储,列包括:feat(二进制格式的 fp16 AudioVAE 潜在表示,形状为 [feat_t, 64])、feat_t(潜在帧数)、text(带语言标签前缀的转录文本)、dataset_id(每种语言的数字标识符)和 split(训练集 train 或开发集 dev)。数据来源于 ghananlpcommunity/ghana-speech 和 ghananlpcommunity/ghana-english-tts-filtered 数据集。该数据集可直接用于 VoxCPM 微调,无需在训练时进行音频解码或 VAE 编码。
VoxCPM Ghana — Precomputed AudioVAE Latents is the exact training-ready data used to fine-tune the model ghananlpcommunity/voxcpm-ghana. It consists of precomputed VoxCPM-0.5B AudioVAE latents (16 kHz) for 42 Ghanaian languages (including Twi split into twi-asante and twi-akuapem) and filtered Ghanaian English, with language-tagged transcripts. The dataset contains 1,756,157 clips, approximately 3,400 hours of audio at 16 kHz. The AudioVAE latents are 64-dimensional with a hop of 640 (approximately 25 frames per second). Data is stored in parquet shards with columns: feat (binary fp16 AudioVAE latent with shape [feat_t, 64]), feat_t (number of latent frames), text (transcript prefixed with the language tag <|lang:CODE|>), dataset_id (per-language numeric identifier), and split (train or dev, with balanced development clips). It is derived from the ghananlpcommunity/ghana-speech and ghananlpcommunity/ghana-english-tts-filtered datasets. This dataset serves as a drop-in replacement for VoxCPM fine-tuning, eliminating the need for audio decoding or VAE encoding during training.




