Mongolian Speech-to-Text Leaderboard
Gemma 4Operating point: google/gemma-4-E2B-it + rookie-systems/gemma-4-audio-asr-mn-adapter (mps)

Gemma 4 — Mongolian Speech-to-Text Benchmark

Real results from the Gemma 4 Batch API (enhanced operating point, language mn) evaluated across 4 Mongolian datasets and 265 audio samples. Every number below is measured — not marketing. WER, speed, and pricing are shown as-is, brutally honest.

Language:MNDiarization:noneSamples:265Run:Aug 17, 2026, 6:18 PMEndpoint:local (transformers + PEFT)
06710069.8%
Accuracy

265/265 samples transcribed · 100% success rate

06710030.2%
Word Error Rate

Lower is better · across 265 samples

Headline metrics
Character Error Rate14.7%
Avg speed factorrealtime multiple
Total speed factor
Avg latency / sample7.3s
Total audio processed1849.3s30.8 min

Pricing

Gemma 4 list pricing for batch transcription. No discounts, no negotiated rates applied — the raw per-minute rate.

Per 1k minutes
$0
batch
Per minute
$0.0000
effective
Per 1k min (these 30.8 min)
$0.00
would cost
Open source?
proprietary

Pricing source: Gemma 4 public pricing. Duudlaga Flow is shown for context only — this page isolates Gemma 4 so the number is not padded by our own product.

WER & CER by dataset

Word and character error rates per dataset. Lower is better — and these are the real Gemma 4 numbers, which are weak on Common Voice 24.

Common Voice 24 (MN)
Source dataset →
WER
32%
CER
14.3%
Shunya Labs Mongolian Speech
Source dataset →
WER
23.4%
CER
11%
Common Voice 20 (MN)
Source dataset →
WER
30.8%
CER
15.3%
Modern Voice
WER
33.1%
CER
16.9%
WERCER

Dataset summary

Aggregate accuracy, speed, and timing for each dataset.

DatasetSourceSamplesSuccessWERCERAccuracySpeedAudio (s)Proc (s)
Common Voice 24 (MN)Hugging Face →5959/5932%14.3%68.0%0.99×330.6s334s
Shunya Labs Mongolian SpeechHugging Face →6060/6023.4%11%76.6%1.07×645.1s604.8s
Common Voice 20 (MN)Hugging Face →5454/5430.8%15.3%69.2%0.84×269.8s322.5s
Modern Voice9292/9233.1%16.9%66.9%1.02×603.8s589.6s

WER vs speed — per sample

Each dot is one audio sample. The sweet spot is the bottom-left (low error, fast). Gemma 4 sits high on error for many Common Voice 24 samples.

Per-sample results

Ground truth shown verbatim in the Expected column. The result column highlights only the words Gemma 4 got wrong, in red — no strikethrough/swap gymnastics, just the mistakes.

265 rows
redwrong word in the resultplaincorrectly transcribedExpected column shown verbatim as ground truth
#AudioSampleDatasetExpected (ground truth)Gemma 4 resultWERCERI/D/S
1btsee_0001Common Voice 24 (MN)Гэхдээ амьсгал хураахаасаа өмнө танд мэдэж байгаагаа хэлье.Гэхдээ амьсгал хураанхаасаа өмнө танд мэдэж байгаагаа хэлье.12.5%1.7%0/0/1
2btsee_0002Common Voice 24 (MN)Надад заяасан аз жаргал гэдэг ердөө гуравхан сарын хугацаатай байсан гэж үү?Надад заяасан газар бол гэдэг ердөө 3 хоногийн хугацаатай байсан гэж үү?33.3%22.7%0/0/4
3btsee_0003Common Voice 24 (MN)Одоо бид өөрсдөө өвчин эмгэгээсээ салахыг хичээцгээе.одоо бид өөрчлөлтөөр өөрийнхөө нэрийг эсээ салахыг хичээгдэхээ71.4%40.4%1/0/4
4btsee_0004Common Voice 24 (MN)Би бол голдуу хээрээр гэр, хэцээр дэр хийж явдаг хүн.гэ бол голдуу хээрээр гэр, хичээлдээ ирэх гэж явдаг хүн40.0%25.5%0/0/4
5btsee_0005Common Voice 24 (MN)Хан хурмаст уурлаж, Болдоггүй Бор өвгөнийг хор луугаараа ниргүүлэхээр явуулжээ.Хамт хүмүүстэй уулзаж, магадгүй бүр угны хоёр рүү гарч ирэхээр явж байна.120.0%55.8%2/0/10
6btsee_0006Common Voice 24 (MN)Алив наашаа ороод ир гээд гэртээ оров.Алив наашаа ороод ир гээд гэртээ оров.0.0%0.0%0/0/0
7btsee_0007Common Voice 24 (MN)Өө өндөр дээдэс таны тухайд би баталж чадахгүй.Ө өндөр дээд таны тухайд би баталгаажуулахгүй.50.0%21.7%0/1/3
8btsee_0008Common Voice 24 (MN)Харин гурав дахь удаагаас эхлэн хүмүүсийг сонирхож эхлэв.Харин гуравдагх удаагаас эхлэх хүмүүсийг сонирхож эхлэв.37.5%7.1%0/1/2
9btsee_0009Common Voice 24 (MN)Та нар очингуутаа шөл л өгч үз.Та нар орчингуутаа шөлө л өгч үз.28.6%6.7%0/0/2
10btsee_0010Common Voice 24 (MN)Ерөөсөө литр үйлдвэрээсээ салаагүй явсан юм чинь.ерөөсөө л үйлдвэрээсээ салавгийн ёс юм чинь42.9%20.8%0/0/3
Page 1 of 27

Methodology

How these numbers were produced.

Provider: Gemma 4 (google/gemma-4-E2B-it with the rookie-systems Mongolian ASR LoRA adapter, run locally on-device (open weights, no API).).

Endpoint local (transformers + PEFT). Language mn. Operating point google/gemma-4-E2B-it + rookie-systems/gemma-4-audio-asr-mn-adapter (mps). Diarization none.

Datasets: Common Voice 24 (MN), Shunya Labs Mongolian Speech, Common Voice 20 (MN), Modern Voice — 265 samples, 1849.3s of audio total.

Dataset source URLs:

Metrics: WER and CER are computed with a standard word/character Levenshtein alignment, normalized for case and punctuation. Accuracy = 100 − WER. Speed factor = audio duration ÷ processing time (× realtime). All requests are real Gemma 4 Batch API calls, not cached or simulated.

Diff highlighting: The result column aligns to the ground truth and colors every substitution and insertion red. Deletions (words missing from the result) are not shown in the result column — the Expected column already holds the full ground truth as-is.

Generated by the Gemma 4 Benchmark Runner · Gemma 4 Batch API v2 · run Aug 17, 2026, 6:18 PM