Choose the right model for every voice task

The Voice Standard Bench scores every model on AssemblyAI’s LLM Gateway on the jobs developers actually run over transcripts: information extraction, summarization, text correction and dictation cleanup. Each model is run on real transcripts at four lengths, from 30 seconds to an hour, and reported on three figures: a quality score from 0 to 100, cost per hour of audio and median response time. The run is read live from the Gateway, so new models appear here as they are added.

Quality is the model’s output scored against each task’s criteria on a 0 to 100 scale, rolled up across tasks for the Overall view. Cost per hour of audio is what the run spent on a model, divided by its calls and scaled to an hour of transcript, at the Gateway’s pass-through provider rates with global routing. Response time is the median across every call at that transcript length, so a single slow call does not move it.

It depends on the task and on how long your transcripts are, which is why this page has a tab for each task and a picker for length. For a given view, the Highest Quality card is the highest score, Lowest cost is the lowest cost per hour of audio and Fastest is the lowest median response time; the scatter shows which models sit on the price-performance frontier. Short transcripts and long ones often favour different models, so set the length before you read the ranking.

Accurate enough to ship for most models on short and medium transcripts, with a wider spread as transcripts get longer. The Summarization tab shows every Gateway model’s score at each length, so you can see where quality holds and where it falls off rather than taking a single headline number. Summaries are generated from a transcript, so Speech-to-Text accuracy sets the ceiling; the benchmark runs on real transcripts for that reason.

Run the models that look right here on your own audio in the LLM Gateway playground: upload a file, run a prompt over the transcript and switch models to see how the answers differ, before writing any code. When you are ready to automate it, the Gateway is OpenAI-compatible, so point your existing SDK at it with your AssemblyAI API key and loop over the model string. The free tier needs no credit card.

On this page it ranges from under a cent to a few dollars per hour of audio depending on the model, the task and the transcript length, and the cost card shows the exact figure for each. The Gateway charges provider rates with 0% markup; the figures shown use global routing, and in-region US or EU routing is 10% higher. You pay per input and output token of the model you pick, with no minimum commitment.