koedesk STT Bench — Keputusan Track A

Ketepatan pertuturan-ke-teks (speech-to-text) berbilang bahasa, diukur menggunakan metodologi yang telah didaftarkan terlebih dahulu — Track A (FLEURS), 13 varian enjin hosted × 12 bahasa × 200 ujaran.

Konflik kepentingan, dinyatakan terus terang di awal: penanda aras (benchmark) ini dibina dan dibiayai oleh koedesk, produk dikte yang menggunakan ElevenLabs Scribe v2. Kami tidak berpura-pura neutral — sebaliknya, segala yang anda perlukan untuk tidak mempercayai kami dan menjalankan semula penanda aras ini sendiri telah diterbitkan: metodologinya (di-commit sebelum larian utama dijalankan — lihat sejarah commit), setiap respons API mentah (results/raw), semua kod pemarkahan, dan audit terma perkhidmatan yang menentukan enjin mana yang boleh disertakan. Di mana Scribe kalah, keputusan itu tetap diterbitkan tanpa diubah.

v1 hanya merangkumi Track A. Track A menggunakan set ujian awam yang mungkin telah digunakan untuk melatih enjin — satu batasan yang diketahui bagi setiap angka pada halaman ini, yang didedahkan dan bukan diselesaikan. Track B (audio rakaman baharu) dibatalkan sebelum sebarang pengukuran dilakukan: rujukan yang dihasilkan oleh manusia tidak serasi dengan reka bentuk penanda aras ini yang sepenuhnya deterministik dan boleh dikira semula (Amendment 7).

Cara membaca jadual

Enjin diberi kedudukan mengikut setiap bahasa ke dalam kumpulan setara (T1, T2, …): sesuatu enjin berkongsi kumpulan yang sama apabila paired bootstrap (n=10,000, seed dikunci semasa preregistration) tidak dapat membezakannya secara statistik daripada peneraju kumpulan tersebut (CI 95% bagi perbezaan berpasangan merangkumi sifar). Perbezaan di dalam satu kumpulan tidak bermakna dari segi statistik — termasuk untuk Scribe v2. Metrik utama ialah CER untuk bahasa Jepun, Mandarin dan Thai, serta WER untuk sembilan bahasa lain; kedua-duanya sentiasa dipaparkan. Lebih rendah lebih baik.

Gambaran keseluruhan — kumpulan mengikut bahasa

Enjin	ja	en	zh	ko	vi	id	th	de	fr	es	pt	ru	#T1
Gemini 3.1 Pro (preview)	T1	T1	T2	T2	T1	T1	T1	T1	T1	T1	T1	T1	10
Microsoft MAI Transcribe (Azure Speech)	T1	T2	T1	T1	T2	T2	T1	T1	T1	T1	T1	T1	9
ElevenLabs Scribe v2 digunakan oleh koedesk	T1	T2	T3	T1	T2	T1	T2	T2	T1	T1	T1	T1	7
Google STT v2 (Chirp 3)	T2	T2	T3	T2	T2	T1	T1	T1	T2	T1	T1	T1	6
OpenAI gpt-4o-transcribe	T2	T1	T2	T2	T3	T2	T2	T1	T1	T1	T1	T1	6
Gemini 3.5 Flash	T2	T2	T3	T2	T2	T1	T1	T2	T2	T1	T2	T2	3
Gemini 2.5 Pro	T2	T3	T6	T2	T2	T1	T2	T3	T2	T2	T2	T1	2
Gemini 2.5 Flash	T4	T4	T6	T3	T4	T3	T2	T3	T4	T2	T3	T1	1
AmiVoice (会話_汎用)	T5	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a	0
Deepgram Nova-3	T4	T4	T5	T6	T5	T5	T3	T4	T4	T3	T4	T4	0
Gemini 3.1 Flash-Lite	T3	T3	T5	T2	T4	T2	T2	T3	T3	T2	T2	T2	0
Mistral Voxtral Mini Transcribe V2	T3	T3	T4	T4	n/a	n/a	n/a	T2	T2	T2	T2	T3	0
OpenAI whisper-1	T4	T3	T5	T5	T5	T4	T4	T3	T3	T2	T3	T3	0

“n/a” = enjin tidak menyokong bahasa tersebut (AmiVoice: bahasa Jepun sahaja; Voxtral: tiada vi/id/th). #T1 mengira bilangan bahasa yang mana enjin berada dalam kumpulan teratas.

Keputusan mengikut bahasa — Track A (FLEURS)

Jepun metrik utama: CER · 200 ujaran

Kumpulan	Enjin	CER% (95% CI)	WER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	1.91 (1.33–2.60)	2.32 (1.83–2.84)	0.22 / 0.30	$0.360	2026-07-05
	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	2.26 (1.75–2.85)	2.69 (2.20–3.20)	0.16 / 0.20	$0.220	2026-07-05
	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	2.49 (1.96–3.06)	2.88 (2.36–3.44)	0.75 / 1.42	—	2026-07-05
T2	Google STT v2 (Chirp 3) chirp_3	2.92 (2.35–3.52)	3.55 (2.92–4.21)	0.37 / 0.43	$0.960	2026-07-05
	OpenAI gpt-4o-transcribe gpt-4o-transcribe	2.96 (2.34–3.64)	3.50 (2.88–4.18)	0.09 / 0.15	$0.360	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	3.43 (2.72–4.19)	3.80 (3.13–4.50)	0.40 / 0.59	—	2026-07-05
	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	3.75 (2.77–5.04)	4.32 (3.41–5.54)	0.44 / 0.67	—	2026-07-05
T3	Mistral Voxtral Mini Transcribe V2 voxtral-mini-2602	4.18 (3.41–5.02)	4.90 (4.20–5.63)	0.17 / 0.22	$0.180	2026-07-05
T3	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	4.53 (3.72–5.39)	5.41 (4.59–6.29)	0.24 / 0.35	—	2026-07-05
T4	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	5.19 (4.06–6.57)	6.43 (5.31–7.81)	0.33 / 0.49	—	2026-07-05
	OpenAI whisper-1 whisper-1	5.49 (4.67–6.36)	6.71 (5.84–7.63)	0.15 / 0.25	$0.360	2026-07-05
	Deepgram Nova-3 nova-3	5.87 (4.91–6.93)	6.70 (5.81–7.63)	0.18 / 0.25	$0.258	2026-07-05
T5	AmiVoice (会話_汎用) -a-general	7.60 (6.46–8.84)	7.96 (6.85–9.15)	0.18 / 0.23	—	2026-07-04, 2026-07-05

Inggeris metrik utama: WER · 200 ujaran

Kumpulan	Enjin	WER% (95% CI)	CER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	OpenAI gpt-4o-transcribe gpt-4o-transcribe	2.18 (1.63–2.83)	1.13 (0.74–1.63)	0.10 / 0.15	$0.360	2026-07-05
T1	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	2.47 (1.91–3.09)	1.40 (1.03–1.82)	0.82 / 1.84	—	2026-07-05
T2	Google STT v2 (Chirp 3) chirp_3	2.68 (2.09–3.31)	1.30 (0.96–1.68)	0.41 / 0.55	$0.960	2026-07-05
	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	2.75 (2.09–3.47)	1.31 (0.94–1.72)	0.28 / 0.43	$0.360	2026-07-05
	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	2.91 (2.30–3.60)	1.55 (1.17–1.99)	0.19 / 0.27	$0.220	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	2.91 (2.29–3.60)	1.45 (1.07–1.87)	0.43 / 0.72	—	2026-07-05
T3	Mistral Voxtral Mini Transcribe V2 voxtral-mini-2602	3.33 (2.68–4.05)	1.45 (1.09–1.84)	0.20 / 0.31	$0.180	2026-07-05
	OpenAI whisper-1 whisper-1	3.64 (2.98–4.36)	1.64 (1.27–2.05)	0.18 / 0.33	$0.360	2026-07-05
	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	3.66 (2.54–5.31)	2.18 (1.24–3.76)	0.51 / 0.89	—	2026-07-05
	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	3.69 (2.94–4.50)	1.75 (1.33–2.22)	0.30 / 0.54	—	2026-07-05
T4	Deepgram Nova-3 nova-3	4.74 (3.99–5.55)	2.29 (1.86–2.75)	0.20 / 0.33	$0.258	2026-07-05
T4	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	4.84 (3.46–6.67)	3.01 (1.72–4.81)	0.38 / 0.62	—	2026-07-05

Mandarin (Cina, ringkas) metrik utama: CER · 200 ujaran

Kumpulan	Enjin	CER% (95% CI)	WER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	5.25 (3.59–7.16)	4.69 (3.67–5.77)	0.23 / 0.36	$0.360	2026-07-05
T2	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	5.74 (4.06–7.64)	5.20 (4.14–6.38)	0.95 / 1.96	—	2026-07-05
T2	OpenAI gpt-4o-transcribe gpt-4o-transcribe	5.86 (4.17–7.76)	5.35 (4.25–6.54)	0.09 / 0.16	$0.360	2026-07-05
T3	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	6.12 (4.38–8.06)	5.75 (4.59–7.01)	0.17 / 0.23	$0.220	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	6.37 (4.69–8.27)	6.29 (5.20–7.46)	0.46 / 0.73	—	2026-07-05
	Google STT v2 (Chirp 3) chirp_3	6.65 (4.88–8.59)	6.71 (5.50–8.00)	0.40 / 0.49	$0.960	2026-07-05
T4	Mistral Voxtral Mini Transcribe V2 voxtral-mini-2602	6.77 (5.01–8.72)	7.47 (6.13–8.92)	0.17 / 0.26	$0.180	2026-07-05
T5	OpenAI whisper-1 whisper-1	7.21 (5.45–9.17)	7.84 (6.55–9.20)	0.17 / 0.29	$0.360	2026-07-05
	Deepgram Nova-3 nova-3	7.47 (5.66–9.49)	8.95 (7.47–10.49)	0.18 / 0.28	$0.258	2026-07-05
	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	7.67 (5.87–9.66)	12.90 (9.89–16.33)	0.26 / 0.39	—	2026-07-05
T6	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	8.68 (6.71–10.94)	10.36 (8.61–12.41)	0.41 / 0.66	—	2026-07-05
T6	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	48.34 (5.18–137.31)	28.64 (5.74–75.51)	0.46 / 0.77	—	2026-07-05

Korea metrik utama: WER · 200 ujaran

Kumpulan	Enjin	WER% (95% CI)	CER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	10.08 (8.09–12.29)	4.93 (3.45–6.65)	0.16 / 0.21	$0.220	2026-07-05
T1	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	10.60 (8.53–12.81)	5.15 (3.67–6.83)	0.21 / 0.31	$0.360	2026-07-05
T2	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	11.46 (9.36–13.74)	5.30 (3.82–7.00)	0.67 / 1.32	—	2026-07-05
	Google STT v2 (Chirp 3) chirp_3	11.49 (9.47–13.64)	5.30 (3.84–6.97)	0.37 / 0.43	$0.960	2026-07-05
	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	11.59 (9.45–13.87)	5.27 (3.80–6.96)	0.43 / 0.79	—	2026-07-05
	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	11.76 (9.69–13.94)	5.45 (3.98–7.14)	0.25 / 0.35	—	2026-07-05
	OpenAI gpt-4o-transcribe gpt-4o-transcribe	12.31 (10.07–14.75)	5.41 (3.89–7.16)	0.09 / 0.14	$0.360	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	12.48 (10.24–14.83)	5.72 (4.20–7.46)	0.37 / 0.64	—	2026-07-05
T3	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	13.07 (10.96–15.30)	5.82 (4.34–7.50)	0.34 / 0.56	—	2026-07-05
T4	Mistral Voxtral Mini Transcribe V2 voxtral-mini-2602	15.11 (12.90–17.42)	6.09 (4.60–7.80)	0.15 / 0.21	$0.180	2026-07-05
T5	OpenAI whisper-1 whisper-1	16.92 (14.56–19.36)	6.62 (5.12–8.34)	0.15 / 0.24	$0.360	2026-07-05
T6	Deepgram Nova-3 nova-3	22.09 (18.80–25.49)	9.27 (7.45–11.27)	0.17 / 0.25	$0.258	2026-07-05

Vietnam metrik utama: WER · 200 ujaran

Kumpulan	Enjin	WER% (95% CI)	CER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	2.11 (1.58–2.70)	1.13 (0.81–1.47)	0.86 / 1.75	—	2026-07-05
T2	Google STT v2 (Chirp 3) chirp_3	2.72 (2.14–3.37)	1.51 (1.17–1.89)	0.38 / 0.44	$0.960	2026-07-05
	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	2.95 (2.34–3.61)	1.58 (1.17–2.05)	0.21 / 0.31	$0.360	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	3.08 (2.38–3.88)	1.71 (1.28–2.18)	0.40 / 0.60	—	2026-07-05
	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	3.11 (2.48–3.79)	1.74 (1.36–2.15)	0.17 / 0.21	$0.220	2026-07-05
	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	3.49 (2.21–5.19)	2.18 (1.03–3.81)	0.46 / 0.82	—	2026-07-05
T3	OpenAI gpt-4o-transcribe gpt-4o-transcribe	3.55 (2.86–4.30)	1.99 (1.56–2.45)	0.09 / 0.12	$0.360	2026-07-05
T4	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	4.28 (3.44–5.19)	2.32 (1.82–2.89)	0.24 / 0.34	—	2026-07-05
T4	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	4.90 (3.93–6.00)	2.86 (2.05–3.96)	0.33 / 0.52	—	2026-07-05
T5	Deepgram Nova-3 nova-3	9.10 (7.64–10.71)	5.10 (4.04–6.39)	0.17 / 0.26	$0.258	2026-07-05
T5	OpenAI whisper-1 whisper-1	9.60 (8.40–10.85)	4.54 (3.91–5.21)	0.17 / 0.27	$0.360	2026-07-05

Indonesia metrik utama: WER · 200 ujaran

Kumpulan	Enjin	WER% (95% CI)	CER% (95% CI)	Kosong	RTF p50/p90	$ / jam audio	Tarikh larian
T1	Google STT v2 (Chirp 3) chirp_3	1.79 (1.29–2.37)	0.58 (0.38–0.83)	0	0.38 / 0.46	$0.960	2026-07-05
	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	2.03 (1.45–2.66)	0.76 (0.48–1.09)	0	0.16 / 0.22	$0.220	2026-07-05
	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	2.05 (1.46–2.68)	0.78 (0.50–1.12)	0	0.64 / 1.09	—	2026-07-05
	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	2.08 (1.49–2.71)	0.68 (0.45–0.93)	0	0.50 / 0.83	—	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	2.31 (1.69–2.98)	0.91 (0.57–1.31)	0	0.40 / 0.59	—	2026-07-05
T2	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	2.44 (1.80–3.10)	0.83 (0.56–1.13)	0	0.26 / 0.42	—	2026-07-05
	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	2.87 (1.75–4.53)	1.33 (0.51–2.81)	0	0.23 / 0.34	$0.360	2026-07-05
	OpenAI gpt-4o-transcribe gpt-4o-transcribe	2.92 (2.24–3.65)	1.12 (0.77–1.53)	0	0.09 / 0.14	$0.360	2026-07-05
T3	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	3.44 (2.66–4.28)	1.03 (0.72–1.38)	0	0.34 / 0.52	—	2026-07-05
T4	OpenAI whisper-1 whisper-1	6.72 (5.63–7.85)	2.04 (1.60–2.50)	0	0.15 / 0.26	$0.360	2026-07-05
T5	Deepgram Nova-3 nova-3	8.31 (7.01–9.68)	2.52 (1.86–3.30)	1	0.17 / 0.26	$0.258	2026-07-05

Thai metrik utama: CER · 200 ujaran

Kumpulan	Enjin	CER% (95% CI)	WER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	3.91 (3.10–4.80)	—	0.23 / 0.37	$0.360	2026-07-05
	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	5.14 (3.73–6.97)	—	1.05 / 1.99	—	2026-07-05
	Google STT v2 (Chirp 3) chirp_3	5.17 (3.73–7.00)	—	0.40 / 0.52	$0.960	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	5.26 (3.74–7.22)	—	0.48 / 0.88	—	2026-07-05
T2	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	5.48 (4.04–7.33)	—	0.18 / 0.27	$0.220	2026-07-05
	OpenAI gpt-4o-transcribe gpt-4o-transcribe	5.55 (4.29–7.07)	—	0.11 / 0.17	$0.360	2026-07-05
	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	5.67 (4.20–7.53)	—	0.37 / 0.67	—	2026-07-05
	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	5.70 (4.19–7.60)	—	0.26 / 0.44	—	2026-07-05
	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	5.94 (4.22–8.11)	—	0.49 / 0.91	—	2026-07-05
T3	Deepgram Nova-3 nova-3	8.39 (6.82–10.34)	—	0.18 / 0.31	$0.258	2026-07-05
T4	OpenAI whisper-1 whisper-1	12.08 (10.64–13.69)	—	0.21 / 0.36	$0.360	2026-07-05

Jerman metrik utama: WER · 200 ujaran

Kumpulan	Enjin	WER% (95% CI)	CER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	OpenAI gpt-4o-transcribe gpt-4o-transcribe	2.21 (1.63–2.85)	0.67 (0.45–0.91)	0.09 / 0.16	$0.360	2026-07-05
	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	2.26 (1.72–2.85)	0.75 (0.52–1.02)	0.59 / 1.13	—	2026-07-05
	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	2.26 (1.67–2.87)	0.70 (0.47–0.96)	0.21 / 0.34	$0.360	2026-07-05
	Google STT v2 (Chirp 3) chirp_3	2.64 (2.07–3.25)	0.72 (0.53–0.93)	0.37 / 0.48	$0.960	2026-07-05
T2	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	2.80 (2.20–3.41)	0.98 (0.67–1.35)	0.17 / 0.27	$0.220	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	2.93 (2.30–3.60)	0.99 (0.70–1.34)	0.34 / 0.63	—	2026-07-05
	Mistral Voxtral Mini Transcribe V2 voxtral-mini-2602	3.20 (2.48–3.98)	0.81 (0.61–1.04)	0.15 / 0.25	$0.180	2026-07-05
T3	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	3.58 (2.80–4.43)	0.95 (0.72–1.21)	0.24 / 0.43	—	2026-07-05
	OpenAI whisper-1 whisper-1	3.87 (3.05–4.73)	1.20 (0.88–1.57)	0.15 / 0.26	$0.360	2026-07-05
	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	3.98 (2.75–5.58)	1.71 (0.71–3.27)	0.41 / 0.72	—	2026-07-05
	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	4.18 (3.29–5.16)	1.06 (0.81–1.34)	0.32 / 0.53	—	2026-07-05
T4	Deepgram Nova-3 nova-3	5.86 (4.75–7.07)	2.57 (1.80–3.51)	0.17 / 0.26	$0.258	2026-07-05

Perancis metrik utama: WER · 200 ujaran

Kumpulan	Enjin	WER% (95% CI)	CER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	3.43 (2.81–4.09)	1.04 (0.83–1.28)	0.25 / 0.37	$0.360	2026-07-05
	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	3.45 (2.74–4.20)	1.19 (0.92–1.47)	0.81 / 1.54	—	2026-07-05
	OpenAI gpt-4o-transcribe gpt-4o-transcribe	3.55 (2.83–4.28)	1.13 (0.88–1.41)	0.10 / 0.15	$0.360	2026-07-05
	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	3.90 (3.20–4.60)	1.40 (1.09–1.72)	0.18 / 0.27	$0.220	2026-07-05
T2	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	4.21 (3.48–4.99)	1.64 (1.32–1.98)	0.45 / 0.80	—	2026-07-05
	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	4.40 (3.56–5.27)	1.42 (1.11–1.75)	0.55 / 1.03	—	2026-07-05
	Google STT v2 (Chirp 3) chirp_3	4.60 (3.72–5.52)	1.45 (1.14–1.77)	0.40 / 0.52	$0.960	2026-07-05
	Mistral Voxtral Mini Transcribe V2 voxtral-mini-2602	4.91 (4.08–5.79)	1.55 (1.26–1.85)	0.17 / 0.27	$0.180	2026-07-05
T3	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	5.24 (4.30–6.23)	1.79 (1.43–2.17)	0.28 / 0.50	—	2026-07-05
T3	OpenAI whisper-1 whisper-1	6.03 (5.13–6.97)	1.97 (1.63–2.33)	0.19 / 0.33	$0.360	2026-07-05
T4	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	6.77 (5.48–8.36)	2.50 (1.81–3.53)	0.39 / 0.65	—	2026-07-05
T4	Deepgram Nova-3 nova-3	7.57 (6.39–8.81)	2.75 (2.26–3.26)	0.19 / 0.30	$0.258	2026-07-05

Sepanyol metrik utama: WER · 200 ujaran

Kumpulan	Enjin	WER% (95% CI)	CER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	OpenAI gpt-4o-transcribe gpt-4o-transcribe	1.53 (1.07–2.05)	0.56 (0.38–0.78)	0.09 / 0.12	$0.360	2026-07-05
	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	1.61 (1.13–2.16)	0.60 (0.41–0.81)	0.64 / 1.02	—	2026-07-05
	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	1.63 (1.12–2.24)	0.61 (0.40–0.85)	0.17 / 0.22	$0.220	2026-07-05
	Google STT v2 (Chirp 3) chirp_3	1.63 (1.21–2.08)	0.52 (0.38–0.69)	0.38 / 0.44	$0.960	2026-07-05
	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	1.67 (1.24–2.12)	0.59 (0.43–0.78)	0.22 / 0.31	$0.360	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	1.76 (1.29–2.28)	0.62 (0.44–0.83)	0.36 / 0.50	—	2026-07-05
T2	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	2.11 (1.42–3.02)	0.97 (0.49–1.83)	0.45 / 0.66	—	2026-07-05
	Mistral Voxtral Mini Transcribe V2 voxtral-mini-2602	2.17 (1.58–2.80)	0.72 (0.50–0.96)	0.16 / 0.22	$0.180	2026-07-05
	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	2.30 (1.76–2.89)	0.74 (0.55–0.95)	0.24 / 0.34	—	2026-07-05
	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	2.34 (1.84–2.87)	0.72 (0.55–0.89)	0.33 / 0.51	—	2026-07-05
	OpenAI whisper-1 whisper-1	2.70 (2.08–3.40)	0.78 (0.59–1.00)	0.15 / 0.26	$0.360	2026-07-05
T3	Deepgram Nova-3 nova-3	3.99 (3.09–5.05)	1.39 (1.04–1.78)	0.17 / 0.25	$0.258	2026-07-05

Portugis (Brazil) metrik utama: WER · 200 ujaran

Kumpulan	Enjin	WER% (95% CI)	CER% (95% CI)	RTF p50/p90	$ / jam audio	Tarikh larian
T1	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	2.34 (1.79–2.93)	1.10 (0.78–1.43)	0.15 / 0.20	$0.220	2026-07-05
	OpenAI gpt-4o-transcribe gpt-4o-transcribe	2.40 (1.81–3.06)	0.83 (0.59–1.10)	0.09 / 0.13	$0.360	2026-07-05
	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	2.53 (1.95–3.14)	0.91 (0.68–1.16)	0.20 / 0.29	$0.360	2026-07-05
	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	2.61 (2.01–3.24)	1.06 (0.76–1.39)	0.68 / 1.13	—	2026-07-05
	Google STT v2 (Chirp 3) chirp_3	2.63 (2.05–3.22)	0.90 (0.68–1.15)	0.37 / 0.44	$0.960	2026-07-05
T2	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	3.02 (2.34–3.73)	0.99 (0.75–1.26)	0.44 / 0.67	—	2026-07-05
	Mistral Voxtral Mini Transcribe V2 voxtral-mini-2602	3.31 (2.69–3.99)	1.08 (0.85–1.33)	0.15 / 0.22	$0.180	2026-07-05
	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	3.41 (2.66–4.20)	1.14 (0.87–1.44)	0.23 / 0.34	—	2026-07-05
	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	3.80 (2.80–5.08)	1.86 (1.08–3.12)	0.34 / 0.53	—	2026-07-05
T3	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	3.92 (3.20–4.70)	1.35 (1.06–1.68)	0.32 / 0.47	—	2026-07-05
T3	OpenAI whisper-1 whisper-1	3.94 (3.19–4.75)	1.26 (1.00–1.54)	0.15 / 0.25	$0.360	2026-07-05
T4	Deepgram Nova-3 nova-3	5.47 (4.52–6.49)	2.12 (1.66–2.61)	0.16 / 0.22	$0.258	2026-07-05

Rusia metrik utama: WER · 200 ujaran

Kumpulan	Enjin	WER% (95% CI)	CER% (95% CI)	Kosong	RTF p50/p90	$ / jam audio	Tarikh larian
T1	Microsoft MAI Transcribe (Azure Speech) mai-transcribe-1.5	2.22 (1.66–2.84)	0.66 (0.44–0.91)	0	0.23 / 0.35	$0.360	2026-07-05
	Google STT v2 (Chirp 3) chirp_3	2.37 (1.56–3.35)	0.90 (0.41–1.67)	1	0.40 / 0.48	$0.960	2026-07-05
	Gemini 3.1 Pro (preview) google-ai-studio/gemini-3.1-pro-preview	2.45 (1.75–3.29)	1.12 (0.61–1.80)	0	0.83 / 1.44	—	2026-07-05
	Gemini 2.5 Pro google-ai-studio/gemini-2.5-pro	2.74 (1.95–3.66)	1.12 (0.59–1.83)	0	0.49 / 0.76	—	2026-07-05
	ElevenLabs Scribe v2 digunakan oleh koedesk scribe_v2	2.77 (2.02–3.63)	1.22 (0.75–1.79)	0	0.18 / 0.24	$0.220	2026-07-05
	Gemini 2.5 Flash google-ai-studio/gemini-2.5-flash	2.82 (2.14–3.59)	0.91 (0.60–1.28)	0	0.37 / 0.59	—	2026-07-05
	OpenAI gpt-4o-transcribe gpt-4o-transcribe	2.85 (1.88–4.04)	0.93 (0.51–1.47)	0	0.10 / 0.14	$0.360	2026-07-05
T2	Gemini 3.5 Flash google-ai-studio/gemini-flash-latest	2.93 (2.24–3.69)	1.17 (0.80–1.60)	0	0.41 / 0.67	—	2026-07-05
T2	Gemini 3.1 Flash-Lite google-ai-studio/gemini-3.1-flash-lite	3.01 (2.22–3.87)	0.90 (0.58–1.29)	0	0.27 / 0.39	—	2026-07-05
T3	Mistral Voxtral Mini Transcribe V2 voxtral-mini-2602	3.72 (2.94–4.55)	1.18 (0.82–1.60)	0	0.17 / 0.24	$0.180	2026-07-05
T3	OpenAI whisper-1 whisper-1	3.80 (3.00–4.66)	1.08 (0.78–1.43)	0	0.17 / 0.28	$0.360	2026-07-05
T4	Deepgram Nova-3 nova-3	4.51 (3.57–5.58)	1.67 (1.17–2.24)	0	0.19 / 0.29	$0.258	2026-07-05

Nota & artifak yang diketahui

Pengulangan semula / audit

Segalanya terdapat dalam repositori: respons API mentah untuk kesemua 28,400 ujaran, normalizer dan kod pemarkahan yang telah dikunci berserta ujian, log pindaan append-only, serta make_scores.py / make_site.py yang menjana semula results/scores.json dan halaman ini secara deterministik — tiada nombor yang disunting secara manual. Menjalankan semula mana-mana sel hanya memerlukan kunci API anda sendiri.