Transcription

DescriptionEvaluates transcription models on multi-lingual, multi-speaker audio with varying levels of background noise across multiple business domains such as software-development, finance, classifieds, food-delivery, and healthcare. The dataset consists of 150 unique audio samples, with each sample being augmented to generate a low and a high noise version.Number of Samples450LanguageEnglish, Hindi, Portuguese, Polish, Afrikaans and DutchProviderToqan (Synthetically generated audio.)Evaluation MethodAccuracy is reported as 1 - WER (Word Error Rate).Data Collection PeriodAugust 2024
Language
The language of the conversation.
Noise
The level of background noise in the audio sample.
Domain
The business domain of the conversation.

Last updated: August 30, 2024

Share this view
#
Model
Provider
Accuracy
No results.

Have a unique use-case you’d like to test?

We want to evaluate how LLMs perform on your specific, real world task. You might discover that a small, open-source model delivers the performance you need at a better cost than proprietary models. We can also add custom filters, enhancing your insights into LLM capabilities. Each time a new model is released, we'll provide you with updated performance results.

Leaderboard

An open-source model beating GPT-4 Turbo on our interactive leaderboard.

Don’t worry, we’ll never spam you.

Please, briefly describe your use case and motivation. We’ll get back to you with details on how we can add your benchmark.