TTS Pretrain Clones 3M: 3 Million Synthesized English Voice-Clone Utterances

2,967,779 clone utterances across 2,971 English speakers, generated by the echo-tts synthesizer. The dataset was created by SynDataLab and last updated on 2026-04 25. It contains WAV audio at 44.1 kHz, stored in Parquet files, with each speaker represented by 10 voice-clone latents and 100 synthesized texts.

Use Cases

• Pre-training text-to-speech models based on the large-scale, speaker-diverse synthetic audio.
• Developing voice cloning systems based on the speaker latents and generated utterances.
• Benchmarking speech synthesis quality based on the structured generation from 100 texts per latent.
• Analyzing speaker identity consistency in synthetic speech based on the multiple latents per speaker.

Strengths

• Contains 2,967,779 audio utterances, providing a large-scale resource.
• Covers 2,971 distinct English speakers, offering significant speaker diversity.
• Audio is high-fidelity with a 44.1 kHz sample rate.
• Generation process is structured, with 10 latents and 100 texts per speaker.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Data is entirely synthetic, which may not fully capture the nuances of natural human speech.
• Description metadata indicates partial coverage for some speaker IDs.

Provenance

Source

SynDataLab via Hugging Face.

Collection Method

Generated by echo-tts synthesizing English text on speaker latents derived from Qwen3-TTS VoiceDesign base speakers.

Time Range

null

Freshness

Last updated 2026-04-25 17:29:36; freshness should be verified.

Geography

null