ElevenLabs Scribe: English transcription test with hard names

👁 118 reads

ElevenLabs Scribe is a speech-to-text model that transcribes recordings in over 90 languages, recognizes speakers, and labels sounds. We tested it in English with a 59-second fictional interview packed with rare names like Cholmondeley and Beauchamp-Marjoribanks, place names, and 12 numbers, then counted content differences in three Scribe settings and in local Whisper. We also tested it on recordings with café, street, and fan noise.

What Scribe can do

• Models: Scribe v2 for files, Scribe v2 Realtime for live speech (latency around 150 ms according to documentation), and Scribe v2 Medical for clinical recordings. The older Scribe v1 is marked as deprecated.
• English: documentation places it in the “Excellent” group, meaning word error rate (WER) up to 5%.
• Features: speaker recognition (up to 32), word-level timestamps, sound descriptions in brackets, keyterms up to 1000 entries, personal data detection and masking, subtitle export, transcript editing with natural language instructions.
• Limits: file up to 5 GB and 10 hours. Instead of a file you can provide a link, including to YouTube or TikTok videos.

Our test: interview with traps

We prepared the recording ourselves, with no real people involved. Two ElevenLabs voices (female Matilda and male Brian) read a fictional interview about the opening of a Beetle Museum in Bicester using the Eleven v4 model: 59 seconds, 133 words, 12 numbers (amounts, percentages, time, road number), and names chosen to trip up transcription: Cholmondeley, Beauchamp-Marjoribanks, Rosalind, Nene, Marylebone.

On 29 September 2026 we ran it through Scribe v2 with English language and speaker diarization in three variants, and for comparison through Whisper large-v3 (faster-whisper, float16 on GPU and int8 on CPU).

The automatic count flagged “Cholmondeley” vs “Chumley” and “kilometres” vs “kilometers” in the plain Scribe run. Whisper also heard “Chumley” and wrote “Beatle” instead of “Beetle”. Review the full differences list in the data: several are spelling variants of rare proper names that a human might accept. Scribe correctly transcribed Beauchamp-Marjoribanks, Bicester, Rosalind Franklin, and Marylebone.

Numbers: words or digits

By default Scribe writes numbers as they sound: “three point four eight million pounds”, “eighteen point five zero pounds”, “nine thirty”. Only the date at the start came out as digits. For notes and subtitles this is inconvenient, so we added the transcript_edit parameter with the instruction:

Write all numbers, amounts, dates, and times as digits,
e.g. 18.50 pounds, 9:30, 1250. Change nothing else.

We got “£3.48 million”, “62%”, “£18.50”, “22 kilometers”, and “9:30”, while the rest of the text remained untouched. The edited text comes in a separate edited_transcript field, alongside the original. It cost 39 credits instead of 30.

Keyterms can help or hurt

Keyterms are meant to help with proper names. We gave Scribe seven terms including “Cholmondeley”, “Quantslant”, “Bicester”, “Beauchamp-Marjoribanks”, “Nene”, “Rosalind”, and “Marylebone”. The result: all names came out correctly, but the automatic count rose to 4 differences because keyterms forced “pounds” and “percent” to be omitted (written as £ and %). Without keyterms, plain Scribe wrote them out. Keyterms are also charged separately: $0.05 per hour.

Background noise

In the Voice Isolator test we prepared a 25-second synthetic monologue (88 words in English) and added fan, street, and café noise, each as loud as the speech. Scribe transcribed all three versions without error. Only when café noise was 5 dB louder than speech did it make 2 mistakes out of 88 words (2.3%). Interestingly, recordings cleaned beforehand with Voice Isolator produced more errors than the raw ones, so do not clean audio before transcription.

Scribe or Whisper

Our take: if your recording can go to the cloud, Scribe wins on accuracy with names, speaker recognition, and speed. Whisper makes sense when material must not leave your computer or you are transcribing hundreds of hours and zero cost matters. A comparison of local models is in the Whisper AI article, and a review of other tools in the transcription guide.

How to transcribe a recording in Scribe

In the app: Speech to Text, upload a file or paste a link, choose English (automatic detection also works, but manual selection is one less risk) and enable speaker recognition. You can download the finished transcript as subtitles, among other formats. Via API (tested 29.09.2026):

curl -X POST https://api.elevenlabs.io/v1/speech-to-text 
  -H "xi-api-key: $ELEVENLABS_API_KEY" 
  -F model_id=scribe_v2 -F language_code=en -F diarize=true 
  -F "transcript_edit=Write all numbers, amounts, dates and times as digits." 
  -F file=@recording.wav

How much does Scribe cost

API response headers showed us the cost directly: the 59-second interview cost 30 credits and $0.0036, with number editing 39 credits and $0.0047. Transcription is also available in the free plan. Plans are described in the how much does ElevenLabs cost article, and other features in the ElevenLabs guide.

Frequently asked questions

Does ElevenLabs Scribe recognize English well?

In our test yes: 133 words of an interview with tricky names and 12 numbers with 2 automatic content differences. Documentation places English in the language group with error rate up to 5%.

How much does transcription in ElevenLabs cost?

Via API $0.22 per hour of audio in Scribe v2, excluding VAT. Our 59-second recording cost 30 credits.

How do you make Scribe write numbers as digits?

Add the transcript_edit parameter with the instruction “Write all numbers, amounts, dates and times as digits”. The corrected version comes in the edited_transcript field, for a small surcharge.

Does Scribe recognize who is speaking?

Yes, when you enable diarize it separates up to 32 speakers. In our interview it attributed all 6 turns to the right people.

Is Scribe better than Whisper?

On our recording yes: Scribe made 2 automatic differences, Whisper large-v3 made 7 on GPU and 6 on CPU, and was 2.2 times slower on GPU and 12 times slower on CPU. Whisper runs locally and costs nothing, though.

Kinetyra

The link to ElevenLabs is a partner link: if you buy through it, QUANTSLANT earns a commission and you pay the same price. It has no influence on the content, including where the coverage is critical.