Leaderboard ng Mongolian speech-to-text

Ihambing ang tasa ng error sa salita, bilis, at presyo ng mga provider ng speech-to-text para sa Mongolian. Bawat numero ay sinusukat sa tunay na Mongolian audio — hindi marketing.

Datasets used

Public speech corpora used for every provider run. Open a source URL for the upstream dataset card, license, and download.

Common Voice 24 (MN)›https://huggingface.co/datasets/btsee/common-voices-24-mn
Shunya Labs Mongolian Speech›https://huggingface.co/datasets/shunyalabs/mongolian-speech-dataset
Common Voice 20 (MN)›https://huggingface.co/datasets/warmestman/common-voice-20-mn-normalized
Modern Voice›Recorded first-party for this benchmark — audio not distributed, metrics only.

WER vs bilis

Bawat tuldok ay isang provider. Ang ideal na posisyon ay sa kaliwang baba (kaunting error, mabilis). Sinusukat sa tunay na Mongolian audio.

Google STTDuudlaga FlowSpeechmaticsChimegeOpenAI WhisperGPT-4o TranscribeAzure SpeechGladiaGemma 4Whisper large-v3Whisper large-v3-turboWhisper large-v2 MNSeamlessM4T v2MMS-1B-allMoonshine MNwav2vec2 XLSR-53 MNWhisper turbo MNWhisper medium MNElevenLabs Scribe v1ElevenLabs Scribe v2Qwen3-ASR-FlashDolphin smallOmniASR CTC-1BOmniASR LLM-1BVibeVoice ASRWhisper large-v3 MN

Paghahambing ng provider

Ayusin ang anumang column. I-click ang Detalye para sa paghihimay per sample, presyo at metodolohiya ng isang provider.

Open source

Provider	WER ↑	CER	Katumpakan	Bilis	Latency	Presyo / 1000 min
Speechmatics	14.4%	7.4%	85.6%	0.51×	13.8s	$8.5	Detalye
Duudlaga FlowAmin	14.6%	7.6%	85.4%	1.47×	5.2s	$19.4	Detalye
Chimege	14.6%	6.6%	85.4%	1.51×	4.7s	$11.1	Detalye
Google STT	16.9%	7.4%	83.1%	1.98×	3.7s	$16	Detalye
Whisper large-v2 MN	20.7%	8.3%	79.3%	0.53×	13.5s	$0	Detalye
Azure Speech	20.8%	9.7%	79.2%	0.39×	17.3s	$16.7	Detalye
Whisper medium MNOverlap	22.5%	8.9%	77.5%	2.7×	2.7s	$0	Detalye
SeamlessM4T v2	24%	11.1%	76%	0.39×	26.3s	$0	Detalye
Whisper turbo MNOverlap	25.7%	9.2%	74.3%	2.43×	3.0s	$0	Detalye
ElevenLabs Scribe v1	27.1%	10.8%	72.9%	2.14×	3.3s	$6.7	Detalye
ElevenLabs Scribe v2	27.1%	10.6%	72.9%	1.51×	4.7s	$3.7	Detalye
Gemma 4	30.2%	14.7%	69.8%	1×	7.3s	$0	Detalye
OmniASR LLM-1B	33.8%	14.2%	66.2%	0.27×	26.6s	$0	Detalye
Whisper large-v3 MNOverlap	39.4%	14.9%	60.6%	1.57×	4.5s	$0	Detalye
MMS-1B-all	41.7%	12.3%	58.3%	19.2×	0.4s	$0	Detalye
wav2vec2 XLSR-53 MNOverlap	45%	16.4%	55%	42.32×	0.2s	$0	Detalye
Dolphin small	49.1%	19.2%	50.9%	1.73×	4.1s	$0	Detalye
OmniASR CTC-1B	51.2%	15.5%	48.8%	0.6×	11.7s	$0	Detalye
GPT-4o Transcribe	53%	27.6%	47%	2.11×	3.7s	$6	Detalye
Moonshine MNOverlap	58.8%	45.2%	41.2%	8.44×	0.8s	$0	Detalye
Whisper large-v3	89.5%	37.4%	10.5%	1.29×	6.0s	$0	Detalye
Gladia	90.1%	38.9%	9.9%	0.75×	9.4s	$10.2	Detalye
Whisper large-v3-turbo	99%	54.4%	1%	0.91×	8.7s	$0	Detalye
Qwen3-ASR-Flash	103.7%	86.4%	-3.7%	2.19×	3.2s	$2.1	Detalye
OpenAI Whisper	105.2%	63.7%	-5.2%	1.5×	4.5s	$6	Detalye
VibeVoice ASR	107%	61%	-7%	0.53×	12.1s	$0	Detalye
Gemini Flash	Paparating na ang benchmark

Lahat ng resulta ay sinusukat sa tunay na Mongolian dataset. Ang bilis ay × real-time (mas mataas = mas mabilis). Mas mababang WER ay mas mainam.

Overlap— Models with this badge were trained on data that overlaps the benchmark corpora (Mongolian Common Voice — 113 of 173 samples — or the public Shunya Labs corpus). Scores on overlapping corpora are inflated by memorization; judge these models by the per-dataset table on their details page, especially the corpora they were NOT trained on.

Duudlaga Voice Set

92 samples · 10.2 min · 28 systems

A second corpus recorded first-party for this benchmark, covering what the public Mongolian datasets barely contain: modern loanwords, English/Mongolian code-switching, numbers, dates, and commands. Every system is measured on the same 92 recordings.

Provider	WER	CER	Katumpakan	Bilis
1Duudlaga FlowAmin	9.5%	5.3%	90.5%	1.27×
2Google STT	11.7%	5.5%	88.3%	2.28×
3Chimege	14.5%	6.8%	85.5%	1.48×
4Speechmatics	15.3%	7.6%	84.7%	0.5×
5Gemini 3.5 Flash	19.5%	11.4%	80.5%	0.06×
6SeamlessM4T v2Open source	22.9%	10.4%	77.1%	4.32×
7Whisper large-v2 MNOpen source	23.2%	10.9%	76.8%	0.55×
8ElevenLabs Scribe v1	23.4%	9.3%	76.6%	2.09×
9ElevenLabs Scribe v2	24.6%	9.8%	75.4%	1.47×
10Azure Speech	26.1%	10.9%	73.9%	0.32×
11Gemma 4Open source	27.7%	13.2%	72.3%	1.04×
12OmniASR LLM-1BOpen source	34.5%	14.5%	65.5%	0.27×
13Whisper medium MNOpen sourceOverlap	35.1%	14.8%	64.9%	2.71×
14Whisper turbo MNOpen sourceOverlap	37.6%	14.1%	62.4%	2.59×
15Whisper large-v3 MNOpen sourceOverlap	41%	17.1%	59%	1.58×
16GPT-4o Transcribe	44.3%	23.7%	55.7%	2.62×
17MMS-1B-allOpen source	45.3%	15.1%	54.7%	20.52×
18Dolphin smallOpen source	46.1%	19%	53.9%	1.69×
19W2v-BERT 2.0 MNOpen sourceOverlap	46.3%	18.2%	53.7%	28.12×
20OmniASR CTC-1BOpen source	54.9%	18.9%	45.1%	0.59×
21wav2vec2 XLSR-53 MNOpen sourceOverlap	56.3%	23.9%	43.7%	45.7×
22Whisper large-v3Open source	87.3%	36.6%	12.7%	1.55×
23Gladia	88.1%	39.9%	11.9%	0.74×
24Moonshine MNOpen sourceOverlap	96.2%	75.7%	3.8%	8.36×
25Whisper large-v3-turboOpen source	98.1%	57.1%	1.9%	1.19×
26Qwen3-ASR-Flash	103.6%	89.9%	-3.6%	2.07×
27OpenAI Whisper	104.6%	63.8%	-4.6%	1.32×
28VibeVoice ASROpen source	110.1%	64%	-10.1%	0.4×

Recorded first-party for this benchmark — audio not distributed, metrics only.