handy models

speech-to-text benchmarks for transcribe.cpp, the engine behind Handy

model detail

Where this model stands against the best model in each language at Q8_0, its speed on an AMD Ryzen 4750U laptop, and what the card says. Add a second model to compare them head to head.

model cohere-transcribe-03-2026
AMD 4750U GPU8.1×
AMD 4750U CPU4.3×
Q8_0 size2.25 GB
frontier · AMD 4750U GPU3 of 14

verdict: Best in 1 of 14 languages, and tied with the leader in 2 more (overlapping 95% CIs). Weakest relative to the field in Japanese (CER) (2.2× the leader's error), English (1.6×), and Mandarin Chinese (CER) (1.6×). Faster than realtime on the AMD 4750U GPU (8.1×) and AMD 4750U CPU (4.3×).

license: Apache-2.0 · card languages: 14 · —

standing by language · 14 benchmarked

Each bar compares this model with the best other model in that language, on a log scale: left of center it is ahead, right of center behind, and the other model is named on the row. The two error rates use the legend colors: this model first, the other model second. Faded bars are ties, where the 95% confidence intervals overlap. Sorted strongest standing first. Hover a row for full details.

cohere-transcribe-03-2026 error best other model error tie: the 95% CIs overlap

quantization · 14 usable languages at Q5_K_M

Q5_K_M is within noise on every one of 14 usable languages: 73% of the Q8_0 file for no measurable accuracy cost.

Q5_K_M file73% of Q8_01.65 GB against 2.25 GB
median error change−0.4%over 14 usable languages
measurably worse0 of 14above the Q8_0 95% confidence interval
worst usable languageItalian3.24 → 3.36 (+0.12 pts)
measurably worse than Q8_0: above its 95% confidence interval within noise typical noise of one measurement
per-language numbers (14)
Italian 3.242.86–3.61 3.36+0.12 +3.7%
Polish 6.155.51–6.85 6.34+0.19 +3.1%
Dutch 7.166.15–8.15 7.23+0.07 +1.0%
Japanese (CER) 5.134.48–5.84 5.17+0.04 +0.8%
Arabic 13.6012.12–15.16 13.62+0.02 +0.1%
French 5.234.72–5.77 5.23+0.00 +0.0%
Greek 8.968.23–9.73 8.94−0.02 −0.2%
English 5.084.56–5.59 5.05−0.03 −0.6%
Vietnamese 7.396.80–7.98 7.31−0.08 −1.1%
Korean (CER) 6.575.50–7.65 6.49−0.08 −1.2%
Mandarin Chinese (CER) 11.1810.19–12.21 11.03−0.15 −1.3%
Portuguese 5.184.71–5.66 5.03−0.15 −2.9%
Spanish 3.973.56–4.39 3.83−0.14 −3.5%
German 5.064.53–5.65 4.79−0.27 −5.3%

speed · all 73 models on the AMD 4750U GPU

× realtime, log scale. Left of the realtime marker, transcribing takes longer than the recording. Hover a dot for the model.

compare with

cohere-transcribe-arabic-07-2026same familywhisper-large-v3ahead of it in 10 of 14 shared languagesVoxtral-Mini-4B-Realtime-2602closest coverage, 11 languages in common