Question · Languages and accents
How accurate is speech recognition in Bulgarian?
On our Mac benchmark, Bulgarian read speech is transcribed with a word error rate of 13.42 percent by Parakeet and 15.87 percent by Whisper large-v3-turbo. That is the weaker end of the field, and Parakeet is slightly ahead.
Updated October 19, 20263 min read13 cited facts, each traced to its source
Bulgarian is usable but needs proofreading: about 13 to 16 wrong words per 100 on read speech. In our October 2026 benchmark on a Mac with an M2 Pro, Parakeet TDT 0.6B v3 had a word error rate of 13.42 percent 1 and Whisper large-v3-turbo had 15.87 percent 2, on 100 Bulgarian recordings from the FLEURS test set. We have no Apple SpeechTranscriber result for Bulgarian.
Parakeet is about two points better than Whisper in Bulgarian
Parakeet has the lower error rate and is faster, at about 18 times faster than real time 3, against about 12 times for Whisper 4. Both are fast enough that you will not wait on either. The speed dataset covers every language.
Our measurement of the engines in Bulgarian
Lower is better in this table.
| Engine | Word error rate (WER %) | Source |
|---|---|---|
| Parakeet TDT v3 | 13.42 | Calculated 1 |
| Whisper large-v3-turbo | 15.87 | Calculated 2 |
The transcripts behind each row are on the on-device benchmark page.
Published results sit within a point of ours
NVIDIA reports 12.64 percent for Parakeet on FLEURS Bulgarian 5, close to our 13.42 percent. OpenAI's paper gives 14.6 percent for the older Whisper large-v2 6, a little better than our turbo result. So the two engines are close, and the order between them is a small effect that could shift with other recordings.
| System | Kind | Test set | Speech | Word error rate (WER %) | Source |
|---|---|---|---|---|---|
| NVIDIA Parakeet TDT 0.6B v3 | Open model | FLEURS | Clean read speech | 12.64 | NVIDIA, 2025 5 |
| Parakeet TDT 0.6B v3 | Open model | FLEURS | Clean read speech | 13.42 | Calculated 1 |
| Whisper large-v2 | Open model | FLEURS | Clean read speech | 14.6 | Radford et al., 2022 6 |
| Whisper large-v3-turbo | Open model | FLEURS | Clean read speech | 15.87 | Calculated 2 |
All published rows are on the published accuracy page.
Bulgarian ranks eighteenth of 25 languages
Taking the best engine in each language, Bulgarian is eighteenth of the 25 in the benchmark. It is far behind Russian (4.59 percent with Whisper 7) and Ukrainian (5.65 percent 8), the other Cyrillic languages. It trails Greek (11.96 percent 9) and is ahead of Latvian (17.27 percent 10). Writing in Cyrillic is not the problem. Bulgarian simply has less training material.
Formatting adds about two points
Counting punctuation and capital letters as a reader sees them, Parakeet's text differed from the written sentences by 15.69 errors per 100 tokens 11, and Whisper's by 17.71 12.
What to expect when you dictate in Bulgarian
In a quiet room, expect roughly 1 wrong word in 7 with Parakeet and a little more with Whisper, so proofread every dictated passage. Errors are most likely on names, numbers, foreign words and word endings. Parakeet is the first engine to try, and it runs entirely on the Mac. Speak in full phrases rather than single words, and keep a correction list if your app offers one. On Windows, Microsoft lists 46 languages for voice typing in Windows 11, from Bulgarian to Vietnamese 13, so Bulgarian is covered there. We found no published fact on Bulgarian in Mac dictation.
Choosing between the two engines matters little in Bulgarian
With two points between them, the choice of engine is mostly about which one your app supports. Parakeet is both more accurate and faster in our test, so it is the default pick, but the published Whisper figure is close enough to ours that you should test your own voice. Dictate one page of your usual text through each engine you can run and count the corrections. The published accuracy page shows how other models compare on similar tests.
What these numbers do not show
The recordings are people reading prepared sentences in a quiet room, scored on one Mac with 100 recordings. Spontaneous speech, noise, regional accents and domain vocabulary are not in the test, and we did not measure Apple's model or Microsoft's voice typing. For the measure itself, see the guide on how accurate speech recognition is.
Our guide to European languages in speech to text helps you choose an engine, and the languages overview covers other languages. For comparison, read the same question for Greek, Ukrainian and Russian.
Evidence
Every number on this page, with where it comes from. Open a fact to see the exact words in its source.
- 1.
- 2.
- 3.
- 4.
- 5.
- 6.
- 7.
- 8.
- 9.
- 10.
- 11.
- 12.
- 13.
The data behind these numbers
- Speech Recognition Accuracy: Word Error Rates8 of this page's figures, 673 data points in all
- On-Device Speech Recognition Benchmark: 25 Languages6 of this page's figures, 56 data points in all
- Speech Recognition Speed: Real-Time Factor Data2 of this page's figures, 66 data points in all
- Dictation Punctuation and Capitalization Accuracy2 of this page's figures, 56 data points in all
Go deeper
Related questions
- Does dictation work for both Brazilian and European Portuguese?
- Does speech recognition understand Swiss German?
- How accurate is speech recognition in Croatian?
- How accurate is speech recognition in Czech?
- How accurate is speech recognition in Danish?
- How accurate is speech recognition in Dutch?
- How accurate is speech recognition in English?
- How accurate is speech recognition in Estonian?