EvalSTT: French Government Speech Corpus for Speech-to-Text Model Evaluation

EvalSTT is a public evaluation corpus for speech-to-text models focusing on French administrative language. Created by the French government's DINUM AI department, it contains official speeches, public addresses, and parliamentary questions. The dataset is published for transparency to document and reproduce the government's model evaluation benchmarks.

Use Cases

• Benchmarking speech-to-text model accuracy based on French administrative speech
• Training domain-specific language models based on official government discourse
• Evaluating model performance on formal parliamentary questions
• Analyzing linguistic patterns in public addresses and official speeches

Strengths

• Created by a government AI department (DINUM) for official evaluations
• Focuses on a specific, formal domain of French administrative language
• Published explicitly for transparency and reproducibility of benchmarks

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 temporal and source bias inherent to its collection of official government materials

Provenance

Source

French government's DINUM AI department

Collection Method

Constitued from official speeches, public addresses, and parliamentary questions

Freshness

Last updated 2026-06-18 18:36:14

Geography

France