Skip to content

Dataset · shb777

Synthetic speech from Gemini Flash 2.0 (284 hours)

shb777/gemini-flash-2.0-speech

An audio dataset with paired texts and voiceovers in two voices from Gemini Flash 2.0 for training TTS and ASR.

What's inside

The set includes 94,512 audio files (47,256 recordings per voice) with English transcriptions, totaling 284.31 hours with an average length of about 10.8 seconds. The speech was generated by the Gemini Flash 2.0 model via the Multimodal Live API in two synthetic voices: male (Puck) and female (Kore). The dataset was used in training the Kokoro TTS and LLaSA 1B models.

What you can do with it

  • Training and fine-tuning text-to-speech (TTS) systems, such as in the Kokoro TTS project
  • Training automatic speech recognition (ASR) models and multimodal audio-text models
Tasks
Text → speech, Automatic speech recognition
Languages
English
Size
10K–100K rows
Downloads this month
277
Likes
61

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("shb777/gemini-flash-2.0-speech")