TTS-German: High-Quality German Speech Dataset for Synthesis and Recognition

TTS-German is a high-quality German speech dataset containing 670,509 audio samples totaling 1,250 hours, derived from the CML-TTS German source. The dataset was processed by datadriven-company, with the last update recorded on March 13, 2026. Audio files are standardized to 24kHz mono WAV format, segmented to a maximum of 12 seconds, and include phoneme transcriptions.

Use Cases

• Train text-to-speech models based on high-quality, loudness-normalized German audio.
• Develop automatic speech recognition systems based on WhisperX-transcribed audio segments.
• Benchmark audio quality using the DNSMOS score filter mentioned in the processing pipeline.
• Research phoneme-level speech synthesis based on the IPA phoneme transcriptions.

Strengths

• Contains 670,509 audio samples, providing a substantial volume of data.
• Totals 1,250 hours of German speech audio.
• Underwent a multi-step quality pipeline including loudness normalization, denoising, and a DNSMOS quality filter.
• Audio is segmented at word boundaries to a maximum of 12 seconds, which is suitable for many model inputs.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The original source, CML-TTS German, may introduce inherent biases in speaker demographics or content.
• Row count is unknown, which may limit suitability assessment for specific tasks.

Provenance

Source

Derived from CML-TTS German.

Collection Method

Processed through a pipeline of standardization, transcription with WhisperX, segmentation, denoising with DeepFilterNet, quality filtering, and phoneme conversion.

Time Range

null

Freshness

Last updated 2026-03 13 21:54:23; freshness should be verified.

Geography

null