Nonverbal TTS Filtered Tokens: Pre-Extracted Audio Codec Tokens for TTS Training

Pre-extracted audio codec tokens for TTS training, containing 6,082 samples totaling 15.6 hours of audio. The dataset was created by author somu9 and was last updated on 2026-05-18. It uses the MOSS-Audio-Tokenizer-Nano codec at a sample rate of 48,000 Hz and a frame rate of 12.5 Hz.

Use Cases

• Training text-to-speech models based on pre-extracted audio codec tokens.
• Fine-tuning speech synthesis systems based on the MOSS-Audio-Tokenizer-Nano codec format.
• Researching audio tokenization and codec performance for TTS applications.
• Developing or benchmarking TTS pipelines that require a frame rate of 12.5 Hz.

Strengths

• Contains 6,082 pre-processed audio samples, providing a substantial starting point for model training.
• Offers 15.6 total hours of audio data, which is a concrete volume for speech synthesis tasks.
• Specifies technical parameters including a 48,000 Hz sample rate and 12.5 Hz frame rate, allowing for precise pipeline integration.
• Provides average sample duration (9.2 seconds) and frames per sample (115.3), aiding in batch size and memory planning.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The license is unknown, which may restrict commercial or research use.
• The dataset's source and collection methodology are not detailed, making bias assessment difficult.

Provenance

Source

huggingface datasets platform, uploaded by author somu9.

Collection Method

Pre-extracted from audio using the MOSS-Audio-Tokenizer-Nano codec.

Freshness

Last updated 2026-05-18 06:09:08; freshness should be verified.