Voxtral 3 vs Whisper, tested: a tie in English, about 4× fewer errors in Vietnamese

Mistral shipped Voxtral 3 on 20 October 2026, a voice stack whose batch transcription model, Voxtral Mini Transcribe 3, lists at USD 0.0035 a minute against USD 0.006 for OpenAI's Whisper and gpt-4o-transcribe. We fed the same two narrations, one English and one Vietnamese, to all three and scored every transcript word by word against the script. In English it was a tie. In Vietnamese, Voxtral made about four times fewer word errors and got 37 of 39 names and terms on our list right, against 16 and 21 for the two OpenAI models.

Tested on 21 October 2026 from a laptop in Vietnam. Voxtral: the Speech to Text playground in Mistral Studio on a new Free account, default settings; the downloaded JSON reports the model as voxtral-mini-transcribe-3. OpenAI: POST /v1/audio/transcriptions with whisper-1 and gpt-4o-transcribe, language set to en or vi. One run per file per model.

At a glance

Key results

• English, a tie: 1.6% word error rate for Voxtral, 1.7% for whisper-1 and 2.1% for gpt-4o-transcribe, after taking numbers out of the score. All three spelled every name and term on our list correctly (29 of 29).
• Vietnamese, a clear gap: 1.7% word error rate for Voxtral against 6.8% for whisper-1 and 6.6% for gpt-4o-transcribe.
• English names inside Vietnamese speech: Voxtral wrote "DeepSeek", "Claude Code" and "Claude Opus" correctly every time. whisper-1 wrote "DeepSig" or "DeepSick" and "Cloud Code"; gpt-4o-transcribe wrote "DeepSec" all 13 times. Listed names and terms right: 37/39, 16/39 and 21/39.
• Cheaper on paper: at list price an hour of audio costs USD 0.21 on Voxtral Mini Transcribe 3 and USD 0.36 on either OpenAI model.
• API not free: the Free plan runs the Studio playground, but API keys stay inactive until you buy credits or pick a paid plan.

Hands-on test

Setup

• Audio: the narration of two of our own videos, so we had the exact script.
  • English: 4 min 24 s, 727 words, read by OpenAI's gpt-4o-mini-tts (voice "onyx"). It explains Claude Skills and names Claude, Claude Code, SKILL.md, Excel and PowerPoint.
  • Vietnamese: 3 min 40 s, 693 words, read by Gemini TTS (voice "Aoede"). It is our test of DeepSeek V4.1 Flash inside Claude Code, so it is full of English product names and terms such as bisect, token and Python.
  • Each file is the scene clips joined with ffmpeg, mono, 16 kHz, 64 kbps MP3.
• Scoring: lower case, punctuation removed, then word error rate (WER) and character error rate (CER) with a plain edit distance. The scripts spell numbers out ("twelve", "mười hai") and every model wrote digits ("12"), which inflated all three scores equally, so the headline figures drop numbers from both sides. Vietnamese also has two accepted spellings for some tones (khoá / khóa); we counted them as the same.
• Names and terms: a fixed list per language (product names plus a few terms such as "token" and "description"), counted in the script and in each transcript.
• Price: the per-minute list prices on Mistral's Voxtral 3 page and OpenAI's pricing page, read on 21 October.

Read this before trusting the numbers. This is clean, single-voice synthetic speech: no noise, no accents, no crosstalk. It is closer to a voice-over or a podcast read than to a meeting recording. And it is slightly tilted toward OpenAI: when we made these videos, our pipeline regenerated any line that an OpenAI transcriber could not match closely enough to the script. Every clip in the test had already passed an OpenAI check, and Voxtral still won in Vietnamese.

English: nothing to choose between them

The few differences were formatting, not hearing. The script spells out "SKILL dot M D" and "claude dot A I"; all three wrote "skill.md" and "claude.ai", which is what you would want, and all three turned "eighth" into "8th". Whisper and gpt-4o-transcribe split "frontend" into "front end"; gpt-4o-transcribe contracted "here is" to "here's". Voxtral's one real miss was "the" heard as "to". For English voice-overs, pick on price, speed or the API you already use.

Vietnamese: the names are where it breaks

The "everything" row is high for all three because the script writes out more than 100 number words ("mười hai" for 12) and every model used digits. Remove those and the gap is plain.

Most of the OpenAI errors were English names said in a Vietnamese sentence. whisper-1 never wrote "DeepSeek" (6 times "DeepSig", 7 times "DeepSick"), turned "Claude" into "Cloud" 6 times, and glued the other 3 "Claude Code" into "ClaudeCode", so not one came out as written. gpt-4o-transcribe handled Claude better (3 of 6 "Claude Code") but wrote "DeepSec" all 13 times, "bisec" for bisect and "mô đồ" for "model". Both also slipped on Vietnamese words: "trỏ" (point) became "chỏ", "cài" (install) became "cái", and "xu" (cent) became "su".

Voxtral's errors were fewer and different: "Haiku" heard as "cu", our brand name "Dunggi" as "Zoomg", "effort" as "F4", and bisect_left / bisect_right glued into one word. For subtitles or notes from Vietnamese tech content, that is the difference between fixing a handful of words and fixing every product name.

What it cost and how we ran it

The Free plan was enough for this test: the playground takes up to 10 files of up to 1 GB, with a "Speaker" label on each line, and lets you download TXT, SRT or JSON. The JSON included usage: 264 seconds of audio and 1,408 output tokens for the English file. The Studio API keys page says "Your organization is on the Free plan. Upgrade to activate your API keys"; the upgrade page offers a Pro plan, a Team plan or pay-as-you-go credits.

At list price, the 8 minutes of audio here would cost about USD 0.03 on Voxtral Mini Transcribe 3 and about USD 0.05 on each OpenAI model. The OpenAI runs did cost real money, about USD 0.10 for all four.

Mistral says its realtime model, Voxtral Mini Transcribe Realtime 3, costs USD 0.008 a minute and covers fewer languages. We did not test it, its TTS (USD 13 per million characters) or the voice agent SDK.

What worked

What worked

• Vietnamese with English product names, which is how most Vietnamese tech content sounds: 1.7% WER and almost every name intact.
• English as good as Whisper, at a lower list price.
• A usable free playground with SRT export, handy for one-off subtitles.

Where it fell short

What didn't work

• No API without paying, even for a small test.
• Short or rare names still go wrong ("Haiku" as "cu").
• Code identifiers lose their underscores.
• We did not find a language setting in the playground, so OpenAI got a language hint and Voxtral did not. Voxtral won anyway, but the setup is not perfectly symmetric.
• We cannot tell you how it does on noisy, real-world speech. This was synthetic narration only.

Our verdict

If you transcribe Vietnamese, or any language where people drop English product names into sentences, Voxtral Mini Transcribe 3 is worth trying first. In our test it made about a quarter of the word errors Whisper made and got almost every name right. For clean English voice-over, all three were close enough that price and the API you already use should decide. Run your own noisy recordings through the free playground before you switch a pipeline.