AI-Terms: 12 Audio Samples for Technical AI Terminology ASR Evaluation

12 audio samples of spoken AI news content comprise this ASR evaluation benchmark created by Trelis in 2026. It provides reference transcriptions and entity annotations specifically for technical AI terminology like model names and benchmarks.

Use Cases

• Calculating entity-level Character Error Rate (CER) using character offsets for company and model names
• Evaluating ASR transcription accuracy for technical AI benchmarks using word-level timestamps
• Testing speech recognition performance on novel technical concepts found in AI news content

Strengths

• Word-level timestamps for precise transcription alignment
• Entity annotations with character offsets for CER analysis
• Specialized focus on 4 categories: companies, models, products, and benchmarks

Limitations

• Extremely small sample size of 12 audio records
• Domain-specific bias toward AI news content

Provenance

Source

Trelis

Collection Method

Annotated spoken AI news content

Freshness

Updated March 2026.

Geography

United States