assemblyai/voice-standard-bench is AssemblyAI's open-source benchmark framework that divides speech AI models into four use cases (dictation, information extraction, summarization, text correction) and records quality, cost, and latency together. The merge commit 2ae44dc of October 29, 2026 (PR #8, "Beam 1.0 and gpt 6.1 terra results") changed only the four result files summary.csv, summary.json, use_cases.html, and use_cases.json under results/canonical/0-1-3/, adding the result rows for Beam 1.0 and gpt-6.1-terra. The change size is 1676 lines, with 1451 insertions and 225 deletions.
The table is split into four buckets — 30s, 120s, 900s, 3600s — and each row contains quality_score, total_cost_usd, cost_per_quality, tokens_per_second, and ttft together. The added Beam 1.0 recorded quality_score 94, total_cost_usd 0.040385, and cost_per_quality 0.0008593 in the 30s bucket, while gpt-6.1-terra recorded 95, 0.046959, and 0.0009886. Placed side by side with the existing rows of the same table, gpt-6.1-sol shows 96 and 0.077908 in the 30s bucket, with cost_per_quality 0.001623, and claude-haiku-4-5-20251001 shows 90 and 0.036948. The 30s quality scores of the two new models are lower than sol and higher than haiku, and their cost_per_quality came out lower than sol's.
Quality scores are assigned separately per use case. The 30s row of Beam 1.0 breaks down as dictation 99, summarization 96, information extraction 91, and text correction 89. In the same commit, the 900s row of qwen3-32B was adjusted from quality_score 61 to 62, and cost_per_quality from 0.001826 to 0.001797.
However, quality_score is a value assigned by claude-opus-5-5, which the judge's column of the table points to, so it is not an independent measurement. As the bucket gets longer, latency and cost rise together, and in the 3600s bucket beam-1.0's error_count is 1 and gpt-6.1-terra's is 2, bringing latency_n down to 29 and 28 respectively. This commit contains no license or scoring-procedure changes, and the changes are limited to the result files. The source is GitHub's AssemblyAI/voice-standard-bench commit 2ae44dc. Therefore, this table is better read as reference material for comparing how quality, cost, and latency trade off under conditions with a fixed response-time budget, and it is not yet the stage at which to conclude which model is superior.