Kartoffelphon-2.5M-de-ger: German Speech Dataset for TTS Models

Approximately 2.5 million audio-text snippets and an estimated 7,000 hours of speech, primarily in German, built as foundation data for Kartoffel TTS models. The dataset is composed mostly of CC / CC-BY based podcast audio. It was created by MultiLlasa and last updated on 2026-06-16.

Use Cases

• Train text-to-speech models based on the large-scale German audio-text pairs.
• Fine-tune speech synthesis systems using the podcast audio data.
• Develop speech recognition models for German based on the transcribed audio snippets.
• Create multilingual speech applications leveraging the retained English segments.

Strengths

• Contains approximately 2.5 million audio-text snippets.
• Offers an estimated 7,000 hours of speech data.
• Primarily focuses on the German language.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.
• Data may reflect source bias inherent to the podcast audio collection.

Provenance

Source

MultiLlasa on Hugging Face.

Collection Method

Composed mostly of CC / CC-BY based podcast audio.

Freshness

Last updated 2026-06-16 21:01:48; freshness should be verified.

Geography

Primarily German language content.