Research
Soniox v5 with Eigen: 44.3% Lower WER Than Baseline on Real World Audio
TL;DR: The largest relative WER reduction in the benchmark's reported pairings was for Soniox v5 on the three-hour mixed English/Hindi call set: 12.81% unenhanced versus 7.14% with Eigen, a 44.3% relative reduction. Soniox also improved on the Hindi Kathbath set. These WER results do not measure latency or agent outcomes.
Soniox v5 is the pairing with the largest reported relative WER reduction on Real World Audio in the July 2026 Eigen benchmark. The result is specific to the samples, model version, and scoring process used.
Results across all four datasets
Dataset | Original WER | With Eigen | Absolute reduction | Relative improvement |
|---|---|---|---|---|
LibriSpeech Clean | 2.16% | 1.56% | 0.60 pts | 27.8% |
VoxPopuli | 7.21% | 6.62% | 0.59 pts | 8.2% |
Kathbath (Hindi) | 22.90% | 18.10% | 4.80 pts | 21.0% |
Real World Audio | 12.81% | 7.14% | 5.67 pts | 44.3% |
Soniox v5 improved with Eigen across every dataset tested, including Hindi. The 44.3 percent relative improvement on Real World Audio is the single largest relative gain recorded anywhere in the benchmark, across all model and dataset combinations.
How Soniox compared with competitor enhancement on real-world audio
System | WER on Real World Audio |
|---|---|
No enhancement | 12.81% |
Krisp VIVA 2.0 (BVC) | 12.02% |
ai-coustics Quail VF 2.2L | 15.64% |
Eigen | 7.14% |
Krisp lowered WER by 0.79 points from the unenhanced baseline on this Soniox pairing, while ai-coustics had higher WER. Eigen reduced WER by 5.67 points from the baseline and had 4.88 points lower WER than Krisp.
What this means for a Soniox-based voice agent
Soniox had the second-lowest unenhanced WER among the four ASRs on this set, yet its WER still decreased from 12.81% to 7.14% with Eigen. Use that as a reason to test enhancement even when the baseline transcript seems relatively good. Include quiet and Hindi cohorts to look for regressions, and measure latency separately: the report does not contain an accuracy-versus-latency comparison.
Related benchmark pages
Eigen, Krisp, and ai-coustics on the same Real World Audio set