AssemblyAI details Voice Standard Bench picks for transcript correction

AssemblyAI said its Voice Standard Bench now offers live picks for transcript correction, a narrow post-processing task meant to clean speech-to-text output without changing what was said. The company’s Oct. 28, 2026 post says the benchmark weighs faithfulness, improvement, latency and cost across four transcript lengths, and warns that the fastest or cheapest model is not necessarily the safest choice for accuracy.

What the benchmark is measuring

Transcript correction, as defined by AssemblyAI, is a constrained rewrite task: remove filler words, false starts and stutters; fix punctuation, capitalization and obvious mis-transcriptions; and preserve every fact, name, number and claim as spoken. The company contrasts that with dictation cleanup, summarization and grammar correction, saying those tasks either change the wording intentionally or fix the speaker rather than the recognizer.

AssemblyAI says Voice Standard Bench evaluates models on real call transcripts with simulated ASR errors, then scores outputs on two dimensions that can conflict with each other: faithfulness and improvement. Faithfulness checks whether the model preserves meaning and factual content, while improvement measures whether it actually makes the transcript more readable without over-editing it. The company says the original transcript is not treated as the answer key because it may contain its own mistakes.

Top picks at four transcript lengths

On the Text correction tab, the company listed different leaders by transcript length as of Oct. 28, 2026. For 30 seconds, GPT-5.6 Sol was the Highest Performance pick with a score of 98, a listed cost of $0.19 per hour of audio and a 1.94-second median response time. For 2 minutes, GPT-6.1 Sol led with a score of 94, $0.11 per hour and a 6.76-second median response time. For 15 minutes, Opus 5 topped the table at 91, with a cost of $0.41 per hour and 14.86 seconds median. For 60 minutes, GLM-5.3 led with a score of 88, at about $0.031 per hour of audio and 14.73 seconds median.

AssemblyAI also showed separate Fastest and Lowest cost cards at each length. The company said these are raw picks with no quality floor, and that the low-cost option can break down on longer transcripts. It cited a 60-minute lowest-cost score of 6 and said the better long-form cost comparison is often the top-scoring model itself, not the absolute cheapest one.

Why the longer runs matter

AssemblyAI said short transcripts give models less context to damage and less material to fix, so cheaper models can come close to the top at 30 seconds. As duration increases, the company said over-editing becomes more costly because there is more text to rewrite, while weaker long-context handling can cause models to miss errors later in the transcript. It argued that this is why transcript correction should be judged differently from shorter voice tasks and why output review remains important, especially where proper nouns, numbers or domain terms matter.

The post also says quality can often be improved upstream before any correction model is used. AssemblyAI pointed to keyterm prompting for names and jargon, text formatting settings for numbers and dates, and choosing the right Speech-to-Text model for the audio and language as ways to reduce the number of errors that later need to be corrected.

How AssemblyAI says to use the results

For teams choosing a model, AssemblyAI recommends starting with the benchmark’s live scores, then testing a shortlist on a small gold set from the user’s own transcripts. It says a good evaluation should count both matching edits and over-corrections, since a fluent transcript can still be wrong if the model changes meanings, proper nouns or numbers. The company also provides a sample API workflow through LLM Gateway, using the same key as Speech-to-Text and a correction prompt intended to preserve meaning while removing noise.

The benchmark launch and methodology were first detailed in AssemblyAI posts dated Oct. 13, 2026, with the correction-specific results updated on Oct. 28, 2026. The company says Voice Standard Bench covers summarization, information extraction, text correction and dictation, but it does not test live conversational-agent turns or tool calling.