SeifElden2342532/TTS_emotional

Dataset Summary

TTS_emotional is a speech dataset built for training and evaluating expressive / emotional text-to-speech (TTS) systems. Each example pairs a short spoken-word audio clip with its transcript, a natural-language description of how the line is delivered, and structured metadata about the voice, speaking style/emotion, and speaker gender. Many of the transcripts are short educational explanations (e.g. "why do we forget things", "how does a compass work"), each read aloud in a distinct emotional style (cheerful, whispering, angry, excited, etc.) by one of several synthetic/labeled voices.

• —Rows: 20.8k total
• —Splits: train (18.7k rows), validation (2.08k rows)
• —Modalities: Audio + Text
• —File format: Parquet (auto-converted; original upload format may differ)

Supported Tasks

• —Text-to-Speech (TTS): training or fine-tuning expressive/emotional TTS models conditioned on style and/or voice identity.
• —Style/Emotion-controllable speech synthesis: using the style and text_description fields as conditioning signals.
• —Speech emotion recognition: using style as a label for classifying delivery/emotion from audio.

Dataset Structure

Data Instances

Each row contains one audio clip (2.6–165 seconds long) with its transcript and metadata. Example (fields abbreviated):

{
  "audio": <audio clip>,
  "text": "Oh, that's a great question! We forget things sometimes because our amazing brain is like a very busy closet that needs to make space...",
  "text_description": "A female speaker communicates the lesson in a cheerful way.",
  "voice_name": "Gacrux",
  "style": "cheerful",
  "gender": "female"
}

Data Fields

Data Splits

Dataset Creation

This is a synthetic dataset: all audio was generated using Google's Gemini TTS, not recorded from human speakers. The transcripts — largely short, simply-worded explanations of everyday science, nature, and general-knowledge questions framed in a teacher/tutor style — were synthesized with Gemini TTS's prebuilt voices (matching the voice_name values, e.g. Gacrux, Enceladus, Puck, Leda, Charon, Orus), each conditioned to be read in a specific emotional/delivery style (the style field) and captioned with a short natural-language description (text_description).

Source Data

The transcripts appear to be short-form educational/explanatory text (either written or model-generated) subsequently synthesized into speech via Gemini TTS. The exact pipeline used to produce the transcripts and select style/voice combinations is not documented by the author.

Annotations

The style, voice_name, gender, and text_description fields are generation metadata — i.e. the conditioning inputs given to Gemini TTS to produce each clip — rather than post-hoc human annotations of naturally occurring speech.

Considerations for Using the Data

• —Synthetic speech, not human recordings: since every clip is generated by Gemini TTS, models trained on this data will inherit that system's prosody, pronunciation patterns, and any artifacts, and may not transfer perfectly to natural human speech.
• —Usage rights depend on Gemini TTS's terms: because the audio is generated output from Google's Gemini TTS, redistribution and downstream use may be subject to Google's API/model terms of service in addition to whatever license the dataset author applies. No license is currently listed on the dataset page — confirm both with the author before commercial use.
• —Class imbalance: with 12 voices and 30 style labels spread across ~20.8k rows, some voice/style/gender combinations are likely underrepresented — check label distributions before training.