Every value on this page is a character error rate in percent, measured by us with the protocol of the parent page. Our two models are the 16-bit versions (fihrIA-ASR 1.7B, 16-bit and fihrIA-ASR 0.6B, 16-bit). The Voxtral column is our own run, which is higher than Voxtral’s published figures; Voxtral does not publish a breakdown.

By subject

ClipsOur 1.7BOur 0.6BVoxtral, our rerunMoulSot
Tourism503.403.919.007.78
Physics304.805.595.985.35
Startups334.995.308.568.87
Finance405.776.308.448.20
History406.188.467.9910.25
Maths396.167.877.9011.57
Business406.637.877.907.88
Life and earth sciences406.2610.9710.3312.88
Engineering459.7213.7412.0116.52
Prism clips. Rows are in the order of our 1.7B model's score.

Our 1.7B model has the lowest error of the four systems in every subject. Its easiest subject is tourism, at 3.40%, and its hardest is engineering, at 9.72%. Engineering is the hardest subject for every system. It is also one of the thinnest topics in our training data: about 9 of the 10,099 hours.

By background noise

ClipsOur 1.7BOur 0.6BVoxtral, our rerunMoulSot
Clean1784.425.847.458.55
A little noise1107.579.449.5011.31
Noisy3611.1317.1215.4217.68
Prism clips. The 33 startup clips carry a different noise label and are left out.

Noise is the largest effect. From clean to noisy clips, character error more than doubles for every system: from 4.42% to 11.13% for our 1.7B model. Our 0.6B model loses the most. It is ahead of our Voxtral run on clean clips (5.84% against 7.45%) and behind it on noisy clips (17.12% against 15.42%).

By speaking pace

ClipsOur 1.7BOur 0.6BVoxtral, our rerunMoulSot
Slow1165.446.838.799.14
Medium1006.138.248.039.97
Fast1087.039.7310.0112.21
Prism clips. The 33 startup clips have no pace label.

Fast speech is harder than slow speech for every system, but the effect is smaller than that of noise: 5.44% to 7.03% for our 1.7B model.

Probe, by clip type

ClipsOur 1.7BOur 0.6BVoxtral, our rerunMoulSot
No label235.105.277.758.23
Ordinary code-switching387.439.5013.8214.81
Tricky code-switching114.744.864.059.02
Technical code-switching914.0915.1319.5626.78
Ordinary and technical code-switching56.217.798.4516.91
Silence520.1720.1724.7827.67
Probe clips. Only the label combinations with at least 5 clips are shown, so the rows do not add up to 109.

These groups are small, so read them as indications. Technical code-switching is the hardest kind of speech for our 1.7B model, at 14.09% on 9 clips.

On the 5 clips labelled silence, both fihrIA-ASR models match at 20.17% CER (27.78% WER), outperforming both Voxtral Mini 4B and MoulSot.