---
title: "Deepgram vs AssemblyAI: live STT accuracy and latency, independent benchmark | Openbenchmarks"
description: "Deepgram vs AssemblyAI speech-to-text, compared on a live rolling 7-day Word Error Rate and Time to Final Segment (TTFS). Independent benchmark data on identical audio. Not vendor claims."
url: https://openbenchmarks.com/speech-to-text-benchmark-by-coval/deepgram-vs-assemblyai
---

[](https://openbenchmarks.com/) 

[← stt benchmark](https://openbenchmarks.com/speech-to-text-benchmark-by-coval) 

[benchmarks](https://openbenchmarks.com/) / [stt by coval](https://openbenchmarks.com/speech-to-text-benchmark-by-coval) / deepgram vs assemblyai 

speech-to-text head-to-head · data by Coval 

# Deepgram vs AssemblyAI: Speech-to-Textlatency & accuracy

Measured head-to-head, not marketing: the verdict is **split**: **Deepgram** is faster on median TTFS (90 ms vs 122 ms), while **AssemblyAI** is more accurate (2.1% WER vs 5.1%). Median TTFS gaps of tens of milliseconds are usually not a pick reason; read p95 TTFS with WER. Independent data, refreshed daily, on identical inputs. Price and whether emails, phones, paths, and IDs survive verbatim are not measured.

\[00\] verdict

## Deepgram or AssemblyAI: which is better?

Each provider's best model on each measured axis, last 7d. Lower is better on both:

| Axis | Deepgram (best model) | AssemblyAI (best model) | Measured winner |
| --- | --- | --- | --- |
| **Speed**: median TTFS | nova-3 · **90 ms** | universal-3.6-pro · **122 ms** | **Deepgram** |
| **Accuracy**: avg WER | nova-3 · **5.1%** | universal-3.6-pro · **2.1%** | **AssemblyAI** |

Not measured here: **price, hallucination on silence, endpointing, timestamps, diarization, long-form drift, rate limits, and whether emails, phones, paths, and IDs survive verbatim**. Treat those as vendor claims until measured. Full leaderboard context: [the independent speech-to-text benchmark →](https://openbenchmarks.com/speech-to-text-benchmark-by-coval) 

\[01\] every model

## Every Deepgram and AssemblyAI model, measured

All benchmarked models from both providers on the same pinned dataset, ranked fastest-first (median TTFS), last 7d:

| Model | Provider | TTFS median | TTFS p95 | WER | Samples |
| --- | --- | --- | --- | --- | --- |
| **nova-3** | Deepgram | **90 ms** | 135 ms | 5.1% | 6,381 |
| **nova-2** | Deepgram | **91 ms** | 144 ms | 6.6% | 6,368 |
| **flux-general-en** | Deepgram | **99 ms** | 164 ms | 5.7% | 6,384 |
| **flux-general-multi** | Deepgram | **101 ms** | 164 ms | 7.4% | 6,384 |
| **universal-3.6-pro** | AssemblyAI | **122 ms** | 266 ms | 2.1% | 6,347 |
| **universal-3.5-pro** | AssemblyAI | **128 ms** | 323 ms | 2.3% | 6,385 |

\[02\] source

## Data by Coval

Coval sells voice-agent evaluation infrastructure. Numbers are a rolling 7d aggregate on a pinned dataset under production-realistic conditions. Openbenchmarks mirrors the results with attribution; methodology and runner are open-source.

synced from Coval 2026-10-11 11:51 UTC · [full STT benchmark →](https://openbenchmarks.com/speech-to-text-benchmark-by-coval) · [methodology →](https://github.com/coval-ai/benchmarks) 

\[03\] faq

## Deepgram vs AssemblyAI: common questions

### Is Deepgram faster than AssemblyAI for speech-to-text?

On the current 7d window, Deepgram's fastest model (nova-3) has a median TTFS of 90 ms, vs 122 ms for AssemblyAI's fastest (universal-3.6-pro), so Deepgram is faster on measured median latency. Distributions matter too: p95 is 135 ms for Deepgram vs 266 ms for AssemblyAI.

### Which is more accurate, Deepgram or AssemblyAI?

By Word Error Rate: Deepgram's best model (nova-3) averages 5.1%, vs 2.1% for AssemblyAI's best (universal-3.6-pro). AssemblyAI leads on measured accuracy. Lower is better; WER is measured on identical inputs under production-realistic conditions.

### Which should I pick, Deepgram or AssemblyAI?

The measured verdict is split: Deepgram is faster on median TTFS, AssemblyAI is more accurate (WER). A few milliseconds of median TTFS is usually not a pick reason; read p95 TTFS and WER together. price, hallucination on silence, endpointing, timestamps, diarization, long-form drift, rate limits, and whether emails, phones, paths, and IDs survive verbatim are not measured here.

### Where does this data come from?

Coval sells voice-agent evaluation infrastructure. Every model runs the same pinned dataset under production-realistic conditions. The table is a rolling 7d aggregate, not a single job; Openbenchmarks mirrors it with attribution. Methodology and runner are open-source (Apache-2.0).