Slovak Speech Recognition (ASR) Benchmark

This is the first dedicated Speech-to-Text (STT) benchmark for Slovak: 39 hours of curated audio across 31 models. Existing resources focus on English or multilingual datasets where Slovak is only a few percent of the data. This benchmark fills that gap, comparing models on Slovak data across different domains.

Task: Automatic Speech Recognition
Metric: WER, CER, DSER
Modality: Speech
Published: June 26, 2026
Updated: November 11, 2026
Models: 31

Ranked by: WER, word error rate

Metric direction: lower is better
Published: June 26, 2026
Number of models: 31 models

The best model, Microsoft MAI-Transcribe-2, makes about 1 word error per 15 reference words.

1
6.54%
2
8.14%
3
9.15%
4
10.94%
5
13.29%
6
14.38%
7
16.22%
8
17.94%
9
18.28%
10
18.35%
11
21.64%
12
21.76%
13
24.33%
14
24.36%
15
24.63%

Measures transcription accuracy at the word level. It is the number of word errors (insertions, deletions, and substitutions) divided by the number of words in the reference transcript. Lower is better.

open weightscommercial API

Click a model name to highlight it across all metrics.

Full results table

Original benchmark results: 31 models, 8 columns. On mobile, scroll horizontally to view the table. Click an underlined column heading to sort the models in the chart above by that column’s values.

Slovak Speech Recognition (ASR) Benchmark: full results table
Model	Category		Insertions	Deletions	Substitutions			RTFx, speed
Microsoft MAI-Transcribe-2	commercial API	6.54%	2.21%	1.10%	3.23%	3.29%	10.52%	—
Google Chirp 3	commercial API	8.14%	2.57%	1.55%	4.01%	4.28%	13.16%	—
ElevenLabs Scribe v2	commercial API	9.15%	3.22%	1.33%	4.59%	5.07%	15.35%	—
OpenAI GPT Realtime Whisper	commercial API	10.94%	1.95%	2.73%	6.25%	5.33%	21.16%	—
Microsoft Azure Speech to Text	commercial API	13.29%	2.62%	3.05%	7.63%	6.04%	26.79%	—
NVIDIA Canary 1B v2	open weights	14.38%	1.67%	4.25%	8.46%	6.88%	17.65%	171.3x
OpenAI Whisper Large v3	open weights	16.22%	2.20%	4.17%	9.84%	7.90%	21.52%	195.6x
Oruk Orukeet	open weights	17.94%	2.29%	4.21%	11.45%	7.90%	39.62%	699.5x
NVIDIA Nemotron 3.5 Transcribe 30B-A4B	open weights	18.28%	2.19%	5.94%	10.15%	10.17%	30.05%	23.7x
OpenAI Whisper Large v3 Turbo	open weights	18.35%	2.74%	4.85%	10.76%	9.11%	27.84%	506.9x
Meta Omnilingual ASR LLM 7B	open weights	21.64%	4.46%	9.95%	7.23%	16.20%	36.90%	25.6x
OpenAI GPT-4o Transcribe	commercial API	21.76%	1.03%	13.27%	7.46%	15.11%	24.80%	—
NVIDIA Parakeet TDT 0.6B v3	open weights	24.33%	1.99%	6.14%	16.20%	10.75%	30.02%	707.7x
Meta Omnilingual ASR CTC 3B	open weights	24.36%	1.09%	14.00%	9.27%	15.88%	61.99%	429.5x
Meta Omnilingual ASR CTC 7B	open weights	24.63%	1.18%	14.94%	8.51%	16.81%	63.35%	225.3x
Meta MMS 1B All	open weights	24.69%	3.13%	7.67%	13.89%	12.40%	62.21%	241.7x
Meta Omnilingual ASR LLM 3B	open weights	25.78%	4.13%	12.72%	8.92%	18.90%	59.14%	27.5x
Meta Omnilingual ASR CTC 1B	open weights	26.79%	1.26%	15.14%	10.38%	17.04%	61.96%	745.0x
Meta Omnilingual ASR LLM 1B	open weights	26.94%	3.86%	14.49%	8.59%	20.32%	57.21%	28.1x
Microsoft VibeVoice ASR	open weights	28.15%	5.01%	3.62%	19.52%	12.29%	34.81%	24.1x
OpenAI Whisper Medium	open weights	29.23%	4.02%	5.90%	19.31%	13.04%	31.82%	228.1x
Meta SeamlessM4T v2 Large	open weights	29.35%	2.17%	16.33%	10.85%	21.54%	31.49%	167.1x
Meta Omnilingual ASR LLM 300M	open weights	29.50%	4.16%	12.96%	12.38%	19.10%	59.64%	28.8x
NVIDIA Nemotron 3.5 ASR Streaming 0.6B	open weights	35.37%	2.83%	8.11%	24.44%	14.14%	55.96%	598.9x
Meta Omnilingual ASR CTC 300M	open weights	37.03%	1.75%	15.63%	19.64%	18.27%	70.30%	1337.1x
ESPnet OWSM-CTC v4 1B	open weights	41.10%	3.27%	7.68%	30.15%	14.94%	63.54%	328.8x
OpenAI Whisper Small	open weights	47.87%	7.49%	6.88%	33.49%	19.27%	44.44%	516.7x
ESPnet OWSM-CTC v3.2 FT 1B	open weights	52.37%	2.98%	11.35%	38.04%	19.94%	76.95%	344.0x
ESPnet OWSM-CTC v3.1 1B	open weights	58.38%	2.91%	12.70%	42.77%	22.08%	82.74%	344.7x
OpenAI Whisper Base	open weights	76.23%	11.66%	7.98%	56.60%	31.45%	134.02%	893.3x
OpenAI Whisper Tiny	open weights	99.69%	23.27%	7.47%	68.96%	47.37%	305.65%	1231.5x

Good to know

Models are evaluated exclusively on Slovak speech data. Detailed data source distribution below.
The results were published on SLAIH on June 26, 2026. Individual models may have been evaluated on different dates.
A dash means the model has no result for that task. It is not a zero.
If a model is missing from the table, no result for it is listed here. Its absence does not mean it performs worse.
These results apply to this benchmark and its metric. They do not measure the model’s overall quality.

Dictation for machines

What does word error rate mean in practice? This simulated transcript of a customer service call (50 words, in Slovak) illustrates the selected model’s measured error rate.

Microsoft MAI-Transcribe-2

Rank 1 of 31

6.54%word error rate (WER)

replaced: 2
omitted: 1
added: 1

Simulated transcript. The numbers of replaced, omitted, and added words are based on the model’s measured error rates, scaled to 50 words. The marked words are chosen at random.

Datasets and methodology

Evaluation data

Dataset	Hours	Source
Parliament (SloPalSpeech)	11.9	Hugging Face
Municipality Hearings	11.7	KInIT
Telco Customer Service	3.8	KInIT & partners
Read Literature	4.5	KInIT
Common Voice SK	2.1	Mozilla Data Collective
TEDxSK and JumpSK Lectures	2.1	KEMT NLP
General Read Speech (Lili)	1.5	Mozilla Data Collective
Fleurs SK	1.5	Hugging Face
Total	38.9

Metric notes

WER (Word Error Rate): Measures transcription accuracy at the word level. It is the number of word errors (insertions, deletions, and substitutions) divided by the number of words in the reference transcript. Lower is better.
Insertions: The share of extra words the model added that have no counterpart in the reference transcript, relative to the number of words in the reference. One of the three components that make up WER. Lower is better.
Deletions: The share of reference words the model missed entirely, relative to the number of words in the reference transcript. One of the three components that make up WER. Lower is better.
Substitutions: The share of reference words the model replaced with an incorrect word, relative to the number of words in the reference transcript. One of the three components that make up WER. Lower is better.
CER (Character Error Rate): Measures transcription accuracy at the character level. It is the number of character errors (insertions, deletions, and substitutions) divided by the number of characters in the reference transcript. More forgiving of small differences like word endings, which is useful for a richly inflected language like Slovak. Lower is better.
DSER (Digit Sequence Error Rate): Measures accuracy in transcribing whole number sequences (phone numbers, account numbers, prices). It is the number of incorrectly transcribed digit sequences divided by the number of digit sequences in the reference transcript. A sequence counts as wrong if any digit in it is wrong. Lower is better.
RTFx (Real-Time Factor): Measures transcription speed as the ratio of audio duration to processing time. An RTFx of 10x means the model transcribes 10 seconds of audio in 1 second. Higher is faster.

Publications

Use the results

Data

Full results table with 31 rows, encoded in UTF-8.

Download CSV

Citation

When you use the results, cite the benchmark and its original publication.

Slovak Speech Recognition (ASR) Benchmark. SLAIH – Slovak Language AI Hub, June 26, 2026. https://slaih.sk/en/benchmarks/speech-to-text