Korean STT Accuracy Tested: CLOVA Speech Recognition vs ElevenLabs Scribe, Same Audio, 24 Runs

Hi, this is Sonetho. ⚡

"For Korean speech recognition, isn't Naver still the one to beat?"
You hear it all the time, but almost nobody has actually measured it.
So we did.
We fed the same 8 Korean audio clips into three engines and compared the output character by character against reference scripts.

📊 Results at a Glance (Character Error Rate, lower is more accurate)

First, credit where credit is due.
On pure Korean, CSR is virtually flawless too.
All three engines tied in the 0% range on news-style reading.
If that's all your audio looks like, any of them will do.

🔬 Total measurements by section (8 trials)

The summary table groups the results by range, so I have laid out all 8 trials here.
Each script was synthesized using two different voices.

💥 The Real Battle Was Mixed-In English

The moment brand names and numbers slipped into Korean sentences, the gap blew wide open.
Here are a few of CSR's actual outputs, copied verbatim.

If your audio mixes in English words, think YouTube subtitles, tech content, or a typical modern meeting recording, that gap translates directly into cleanup work.
Scribe preserved ChatGPT, Gemini, and API exactly as written across the same passages.

🎧 The Audio and Transcriptions Used

Numbers alone don't tell the whole story.
I am sharing the raw audio used for the English-mixed sections below.
Please listen along and compare them with the transcriptions.

Audio generated by CLOVA Ara

Ground-truth script
일레븐랩스의 Creator 플랜은 월 22달러지만, 첫 달은 50% 할인된 11달러에 시작할 수 있습니다. ChatGPT나 Gemini 같은 AI 툴과 연동하면 활용 폭이 넓어지고, 2026년 7월 기준 API 요금은 100만 자당 약 15달러 수준입니다. 자세한 내용은 sonetho.com에서 확인하세요.

CLOVA CSR (CER 36.9%)
11 x의 크리에이터 플랜은 월 22 달라지는 첫달은 50% 할인된 11달러에 시작할 수 있습니다 채취 pt 나재민이 같은 ai 툴과 연동 하면 활용 폭이 넓어지고 2026년 7월 기준 api 요금은 100만장 약 시부터 여수 중입니다 자세한 내용은 선 투 닷컴에서 확인하세요

Scribe v1 (CER 13.8%)
일레븐웹스의 크리에이터 플랜은 월 $22지만 첫 달은 50% 할인된 $11에 시작할 수 있습니다. ChatGPT나 Gemini 같은 AI 툴과 연동하면 활용 폭이 넓어지고 2026년 7월 기준 API 요금은 100만 자당 약 $15 수준입니다. 자세한 내용은 son토.com에서 확인하세요.

Scribe v2 (CER 20.8%)
일레븐랩스의 크리에이터 플랜은 월 22달러지만 첫 달은 50% 할인된 11달러에 시작할 수 있습니다. 챗GPT나 제미니 같은 AI 툴과 연동하면 활용 폭이 넓어지고 2026년 7월 기준 API 요금은 100만 자당 약 15달러 수준입니다. 자세한 내용은 소토닷컴에서 확인하세요.

There is an important point to note here.
Although v2 has a worse score than v1, it actually performed better in comprehension.
v1 got the brand name wrong as "일레븐웹스", while v2 got it right as "일레븐랩스".
Regarding the price, v1 converted the symbol to $22, but v2 transcribed it as "22달러" exactly as in the original.
However, v2 phonetically transliterated ChatGPT as "챗GPT" and Gemini as "제미니" in Korean.
These were all marked as errors in the Character Error Rate (CER) calculation, leading to the 20.8% result.
It would be a mistake to conclude that v1 is superior based solely on the scores.

🎙️ Same sentence, yet 4x the error rate depending on the voice

One result was entirely unexpected.
The ground-truth script for the long-form segment is identical.
However, the CSR error rate was 7.3% for the audio read by the CLOVA voice, while it was 1.7% for the version read by the ElevenLabs voice.
The voice itself, rather than the recognizer, caused a fourfold difference in accuracy.
Scribe recorded 0% error for both audio clips.

Long-form segment read by CLOVA Ara

Long-form segment read by ElevenLabs Hyuk

The failure occurred at the end of the sentence.
The ground-truth script is "어제보다 오늘 더 자연스러워지고 있다는 것만은 분명합니다".
For the CLOVA voice, the CSR transcribed it as "있다는것만 있는 영화".
At the same position, the CSR for the ElevenLabs voice accurately transcribed it as "있다는것만은 분명합니다".
Additionally, the "있습니다" in the middle of "근본적으로 바꾸고 있습니다" was omitted entirely only in the CLOVA audio.

Here is what this means for real-world applications.
When measuring STT accuracy, testing only the recognition engine gives you an incomplete picture.
The pronunciation and pacing of the input audio influence the results significantly.
When selecting internal tools, you must perform your tests using the voices of the actual speakers you intend to use.

🧐 Scribe Wasn't Perfect Either

In the interest of honesty, here are Scribe's mistakes too.
It struggles with proper nouns it has never seen.
It wrote "sonetho.com" as "sonto.com", and in one sample it misheard "11달러" ($11) as "1달러" ($1).
If your document has money riding on its numbers, proofreading is non-negotiable with any engine.

The difference between v1 and v2 was interesting as well.
v2 automatically normalizes "22달러" ($22 spelled out in Korean) into "$22".
Our scoring method (CER) counts that as an error, but the meaning is spot on.
If verbatim fidelity matters, go with v1; if you want readable subtitles and documents, v2 is the right pick.
v2 also comes in a real-time streaming version.

💰 How the Pricing Compares

ElevenLabs Scribe: $0.22 per hour (batch).
Even the free plan includes 4.5 hours per month.
The $6 Starter includes 27 hours and the $22 Creator includes 100, so most personal workloads fit comfortably inside a plan.

CLOVA Speech Recognition (CSR): pay-as-you-go, billed in 15-second increments rounded up.
Check the exact rate in the Naver Cloud pricing calculator.
Note that CSR is an API for short utterances with a 60-second limit.
For long recordings and speaker diarization, you need a separate product called CLOVA Speech.

🤔 So Which One Should You Use

When CLOVA CSR is the right call
Short utterances that are mostly pure Korean.
Projects with Korean domestic cloud infrastructure requirements.
Teams already invested in the Naver Cloud ecosystem.

When ElevenLabs Scribe is the right call
Subtitles and meeting notes for content that mixes English and Korean.
IT and marketing audio full of brand names.
You can test the quality yourself on the free 4.5 hours before committing.

🧪 Methodology (Reproducible)

We wrote scripts across 4 categories (emotional dialogue, news style, mixed English and numbers, long form) and synthesized them with two high-quality TTS voices (CLOVA Ara Pro and ElevenLabs Hyuk) to create 8 audio clips.
Each clip went through all three engines and was compared character by character (CER) against the reference script.
Full disclosure on the limitations.
The input is TTS-synthesized audio, so conditions are noise-free and ideal; real human speech or noisy environments may shift the numbers.
Some errors may also stem from the TTS pronunciation itself.

❓ Frequently Asked Questions (FAQ)

Q. Which STT is the most accurate for Korean?
In this test, Scribe v1 posted the lowest average character error rate at 3.2%.
That said, on formal pure Korean, CLOVA CSR also tied in the 0% range, so whether your audio contains English is the real deciding factor.

Q. What's the difference between CLOVA Speech Recognition (CSR) and CLOVA Speech?
CSR is an API for short utterances under 60 seconds, while CLOVA Speech is a separate product that supports long recordings and speaker diarization.
This test covers CSR.

Q. Should I use Scribe v1 or v2?
Pick v1 if preserving the original notation matters, and v2 if you want clean, readable output where things like currency symbols get normalized automatically.
v2 also supports real-time streaming ($0.39 per hour).

For hands-on interview transcription and speaker diarization, the story continues in ElevenLabs vs Whisper vs Deepgram: STT compared.
See you in the next post. This was Sonetho. ⚡