Tag: benchmarks
All the articles with the tag "benchmarks".
Benchmarking Nemotron on legal files
Nemotron found 17% of legal terms on our deposition bench at first. Most of that was number formatting. How it compares with AssemblyAI and WhisperX.
13.1% normalized WER: words before punctuation
One Record Test Bench run scored 13.1% normalized proofread WER. Removing case and punctuation shrinks the measured error. It does not fix the words.
Gemini, Sol and Sonnet: speaker tracking is still not fixed
Speaker tracking is our users' biggest complaint. Compare the published speaker-name scores for Gemini, GPT-6.1 Sol and Claude Sonnet 5.5.
The Record Test Bench is live: first results
The Record Test Bench is live. Audio full-chain results: legal-structure F1 53.7%, speaker-name accuracy 63.9%, best verbatim WER 20.2%.