The Transcribe Labs BlogHow we test transcription quality, and what we found.https://transcribe.solutions/blog2026-09-22T00:00:00ZWhy one engine isn't enough: more engines only help if they fail differentlyhttps://transcribe.solutions/blog/engines-that-fail-differently2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiOne engine gets names and numbers consistently wrong. More engines only help when they fail in different places. A real interview, including where we lost.The most dangerous transcription errors read perfectlyhttps://transcribe.solutions/blog/fluent-errors2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiA flipped negation reads perfectly and a majority vote won't catch it. So tools should show where they're unsure. We checked 15 of them.WER lies. Here's how we measure accuracy on real interviewshttps://transcribe.solutions/blog/wer-lies2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiThe industry's favourite error metric misleads on interview audio. What we use instead: check the reference first, then count errors that change meaning.When interviews switch languageshttps://transcribe.solutions/blog/mixed-language-interviews2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiSpanish interviewee, Chinese researcher, an interpreter in between. Vendor docs and lab tests misled us; a real recording gave the opposite answer.Who said that? Speaker attribution when many people talkhttps://transcribe.solutions/blog/who-said-that2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiBuilt-in speaker labels drift as people are added. Learning each voice first got us above 99% on public meeting data, as long as people take turns.Field notes from 27 languages: how we picked the engines for eachhttps://transcribe.solutions/blog/27-languages2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiNo engine is best everywhere. The index to our language-by-language tests: which engines we picked for each of 27 languages, and why.Arabic: people speak dialect, subtitles write MSA. How we picked engines anywayhttps://transcribe.solutions/blog/languages-ar2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiSame engine, different reference transcript: error rate went from 49% to 18%. Arabic diglossia forced us to judge engines by meaning, sentence by sentence.Czech: the two-letter prefix that decided our engine lineuphttps://transcribe.solutions/blog/languages-cs2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiIn Czech, "not" is often just ne- on the front of a verb. In our hardest clip, Speechmatics lost it and Soniox kept it. How we picked Czech engines.Danish: we swapped out the cleanest engine on the most interview-like cliphttps://transcribe.solutions/blog/languages-da2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiOn the Danish clip closest to a real interview, AssemblyAI was the cleanest engine. We replaced it with Soniox anyway. Our reasoning, and the case against it.German: when Viennese dialect made the steadiest engine switch to Englishhttps://transcribe.solutions/blog/languages-de2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiThree German films and one Austrian. AssemblyAI was cleanest on standard German, then switched to English on Viennese dialect. Doubao never left German.Greek: lose the accent marks and half the words fall aparthttps://transcribe.solutions/blog/languages-el2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiOne engine produced real Greek but swallowed every accented vowel whole. Greek error rates run high by nature. How we picked engines on four film clips.English: two reference tracks that fail in different placeshttps://transcribe.solutions/blog/languages-en2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour English film scenes read line by line. AssemblyAI and Doubao rarely make the same mistake, which is why both back up ElevenLabs.Spanish: the engine our own code made look brokenhttps://transcribe.solutions/blog/languages-es2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour film scenes from Spain, Mexico and Argentina. Doubao was once written off as broken in Spanish. The real cause was a language code we got wrong.Finnish: near-perfect on clear speech, invented lines under gunfirehttps://transcribe.solutions/blog/languages-fi2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiOn clear Finnish conversation, ElevenLabs made no real errors. Under war-film gunfire, every engine invented lines. And why Finnish is where we skip Soniox.French: three engines, one sentence, the same wrong meaninghttps://transcribe.solutions/blog/languages-fr2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour film scenes from the Paris banlieue, Paris and Quebec. Doubao was the only non-main engine that never invented or flipped a line, and WER told us nothing.Hungarian: why we kept the engine that wrote "ticket" 23 timeshttps://transcribe.solutions/blog/languages-hu2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFive Hungarian film clips. AssemblyAI looped one word 23 times in a brawl, then stayed clean at a snack bar. We kept it. Here is the evidence against that call.Indonesian: someone screams for help, the engine writes 'thank you'https://transcribe.solutions/blog/languages-id2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiAll four Indonesian test clips come from one film. What engines invented in the noisy scenes, and why Soniox replaced AssemblyAI as a reference.Italian: when our main track heard dismay as praisehttps://transcribe.solutions/blog/languages-it2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour film scenes from Rome, Sicily, Naples and Milan. On one key line, Doubao was the only engine of ten to get it right; ElevenLabs flipped the meaning.Japanese: the engine that looked best was deleting the fillershttps://transcribe.solutions/blog/languages-ja2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiTwo 1950s films and two modern interviews: how filler words flipped our Japanese ranking, why Soniox replaced AssemblyAI, and what is still unresolved.Korean: three engines, one identical wrong negativehttps://transcribe.solutions/blog/languages-ko2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour Korean film scenes: why ElevenLabs led by a wide margin, why Doubao and Soniox replaced AssemblyAI, and what a three-way shared error taught us.Malay: the speaker switched to English, and the engine translated it backhttps://transcribe.solutions/blog/languages-ms2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiMalay interviews switch between Malay and English mid-sentence. The riskiest error: an engine quietly rewriting the English as fluent Malay.Dutch: no engine crossed a red line. So how do you choose?https://transcribe.solutions/blog/languages-nl2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour Dutch clips, four engines, no invented sentences, flips or loops. When red lines can't separate engines, we look at whose mistakes differ.Norwegian: "we weren't unfaithful" became "we chose to be unfaithful"https://transcribe.solutions/blog/languages-no2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFive Norwegian film clips and a podcast. One engine turned "we weren't unfaithful" into "we chose to be unfaithful" three times, on clear audio. How we chose.Polish: drop one "nie" and "I won't have kids" becomes "I will"https://transcribe.solutions/blog/languages-pl2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour Polish film clips, one flipped negation, and that was enough to keep an engine out. How we picked engines for Polish, and why we swapped one.Portuguese: three flipped lines, and a key sentence rescuedhttps://transcribe.solutions/blog/languages-pt2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiBrazilian film scenes plus three real conversations: Gemini flipped a line in each, yet often beat the main track. European Portuguese has no test clip.Romanian: "Are you hungry?" became "I'm starving"https://transcribe.solutions/blog/languages-ro2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiOn the clearest Romanian clip, one engine flipped three meanings in a minute. Romanian was also the first language where results made us swap a track.Russian: the line both reference engines got wronghttps://transcribe.solutions/blog/languages-ru2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiHow we tested eight speech engines on four hard Russian film scenes, what went wrong, and why we run ElevenLabs with AssemblyAI, Doubao and Gemini.Swedish: when "apologize!" became "I apologize"https://transcribe.solutions/blog/languages-sv2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour Swedish test clips, nine engines. In one scene a bully orders his victim to apologize; four engines turned it into "I apologize." Only one got it right.Thai: part of our primary engine's score was our own bughttps://transcribe.solutions/blog/languages-th2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiThree Thai clips: a bug that stripped vowel and tone marks, Gladia's runaway repeats, why best-scoring Qwen isn't used, and what we still don't know.Turkish: how one model upgrade broke a clean recordhttps://transcribe.solutions/blog/languages-tr2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour Turkish film scenes: why AssemblyAI's newer model did worse, why last-placed Doubao stays, why Soniox is only a backup, and Gemini's missing endings.Ukrainian: 54 seconds of silence, turned into a lecturehttps://transcribe.solutions/blog/languages-uk2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiA stretch with only a TV in the background, and two engines wrote dialogue into it. Why we turned down Soniox for Ukrainian, and the risk we kept.Vietnamese: the 'hallucination' was a market vendorhttps://transcribe.solutions/blog/languages-vi2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiWe have only two valid Vietnamese test clips. What they show, why we run ElevenLabs with Doubao, Soniox and Gemini, and what two clips can't prove.Mandarin: choosing engines through gunfire and screaminghttps://transcribe.solutions/blog/languages-zh2026-09-22T00:00:00Z2026-09-22T00:00:00ZZeyu SiFour Mandarin film scenes, twelve engines, read line by line: who invents sentences in noise, who gives up, and why Chinese ended up with four reference tracks.