How accurate is AssemblyAI Universal 3.5 Pro?

AssemblyAI Universal 3.5 Pro transcribes the 1,000 public Pipecat clips at 1.93% word error rate and the eight licensed Ocular recordings at 4.23%. Word error rate counts incorrect, missing, and extra words against the reference transcript, so lower is better.

How fast is AssemblyAI Universal 3.5 Pro in a live conversation?

First text arrives after 489ms at the median, and the transcript is final 180ms after the speaker stops. Both are measured on the same 206 streamed turns as every other model, and the page shows P90 and P95 as well, which is what a caller meets on a bad turn.

How was AssemblyAI Universal 3.5 Pro measured?

Every model runs the same two datasets under the same conditions: 1,000 public clips from Pipecat's STT benchmark dataset for accuracy at scale, and eight licensed conversation recordings from Ocular for real speech. Latency comes from 206 streamed turns, using each provider's own supported finalization contract. AssemblyAI Universal 3.5 Pro returned usable text for 1,000 of the public clips.

Bench

Book a demo

AssemblyAI Universal 3.5 Pro

Every modelAssemblyAI

Accuracy vs. speed

Lower and further left is better · Nemotron 3.5 Transcribe doesn't support streaming, so it is not shown

Best quadrantBest picks

Final-text delay P50 (ms)

Overall WER (%)

AssemblyAI Universal 3.5 Pro is 2nd of 16 on time to first text and 12th of 19 on the Ocular Dataset.

2.72% Overall WER (7th) · 1.93% Pipecat WER (4th) · 4.23% Ocular WER (12th) · 489ms First text (2nd) · 180ms Final text (10th)Likely range, 95% confidence: Pipecat WER 1.69 to 2.18%, Ocular WER 3.57 to 4.68%Published price $0.45 per hour of streamed audio. AssemblyAI's standard real-time rate, billed while the session is open. Source, checked Oct 8, 2026.

Last updated Sep 16, 2026Methodology by Dileep Chagam

AssemblyAI Universal 3.5 Pro against every other model

Lower is better on every measure. Hover a column for that model's full numbers.

NVIDIA Nemotron 3.5 Transcribe and GPT-4o Mini Transcribe and GPT-4o Transcribe send no text until the speaker stops, so TTFT is not comparable.

Glossary: WER, TTFT, TTFS, P50

WER · Word error rate: Share of words the model got wrong: replaced, missing or added. Lower is better.
Pipecat Dataset: 1,000 short public clips from Pipecat's STT benchmark dataset.
Ocular Dataset: 8 recordings from 4 real phone conversations, licensed from Ocular.
Likely range · 95% range: Where the model's true score probably falls, given how much audio was tested. Overlapping ranges are too close to call.
TTFT · Time to first text: How long after the audio starts streaming until the first words appear.
TTFS · Final-text delay: How long after the caller stops until the full transcript arrives. A voice agent waits for this before it replies.
P50 / P90 / P95 · Typical, slow and slowest turns: Half of turns are faster than P50. Only 1 in 10 is slower than P90, and 1 in 20 slower than P95.

Where AssemblyAI Universal 3.5 Pro struggled

On real calls, dropped words are its biggest source of error. On clean clips, wrong words lead. For costs, see our breakdown of AssemblyAI pricing.

61%

of its errors on the Ocular calls are words it left out.

Median across all 19 models: 49%

Mix of errors

Real calls (Ocular Dataset)

Clean clips (Pipecat Dataset)

TTFT and TTFS at P50, P90 and P95

Timed on the same 206 streamed Ocular turns as every other model. P90 and P95 show what a caller meets on a slow turn.

TTFS, final text

How long after the caller stops until the full transcript arrives. A voice agent waits for this before it replies.

20ms50ms100ms200ms500ms1s2s

Typical, P50: 180ms; 10th fastest of 18
Slow, P90: 532ms; 13th fastest of 18
Slowest, P95: 692ms; 13th fastest of 18

TTFT, first text

How long after the audio starts streaming until the first words appear.

200ms500ms1s2s5s10s20s

Typical, P50: 489ms; 2nd fastest of 16
Slow, P90: 966ms; Fastest of 16
Slowest, P95: 1.08s; 3rd fastest of 16

Every other model's typical turn, on the same axis.

1,000 Pipecat clips · 8 Ocular recordings · 205 turns timed · 1 failed requests

AssemblyAI Universal 3.5 Pro FAQ

Where does AssemblyAI Universal 3.5 Pro place in the benchmark?

AssemblyAI Universal 3.5 Pro places highest on time to first text, 2nd of 16 at 489ms. Its weakest placing is the Ocular Dataset, 12th of 19 at 4.23%.

AssemblyAI Universal 3.5 Pro places highest on time to first text, 2nd of 16 at 489ms. Its weakest placing is the Ocular Dataset, 12th of 19 at 4.23%.
AssemblyAI Universal 3.5 Pro transcribes the 1,000 public Pipecat clips at 1.93% word error rate and the eight licensed Ocular recordings at 4.23%. Word error rate counts incorrect, missing, and extra words against the reference transcript, so lower is better.
First text arrives after 489ms at the median, and the transcript is final 180ms after the speaker stops. Both are measured on the same 206 streamed turns as every other model, and the page shows P90 and P95 as well, which is what a caller meets on a bad turn.
Every model runs the same two datasets under the same conditions: 1,000 public clips from Pipecat's STT benchmark dataset for accuracy at scale, and eight licensed conversation recordings from Ocular for real speech. Latency comes from 206 streamed turns, using each provider's own supported finalization contract. AssemblyAI Universal 3.5 Pro returned usable text for 1,000 of the public clips.

Building on AssemblyAI? Test your agent the same way.

Cekura runs these simulated calls on your agent and scores every one.

Book a demo Start free