Munch 1 Latent New Parquet: Precomputed Urdu TTS Latent Representations

51,021 pre-computed latent representations for Urdu utterances, designed to bypass audio decoding during TTS model training. The latents are derived from the Humair332/Urdu-munch-1 audio source using the Aratako/Semantic-DACVAE-Japanese-32dim codec at a 25 Hz frame rate. Author zuhri025 uploaded this dataset to Hugging Face in April 2026.

Use Cases

• Train Urdu text-to-speech models based on pre-computed latent representations.
• Benchmark TTS model performance using a standardized latent feature set.
• Experiment with latent space manipulation for speech synthesis based on the described codec features.

Strengths

• Contains 51,021 pre-processed Urdu utterance latents, providing a substantial starting point for model training.
• Specifies technical parameters like a codec sample rate of 48,000 Hz and a latent frame rate of 25.0 Hz.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect source bias inherent to the original Humair332/Urdu-munch-1 audio collection.

Provenance

Source

Humair332/Urdu-munch-1

Collection Method

Pre-computed DACVAE latent representations.

Freshness

Last updated 2026-04-18 16:28:59; freshness should be verified.