Audio language is independent of this page’s language. These subtitles are in the spoken language, not translations.
Audio language: Deutsch · de_de
de-01 · Deutsch · de_de
First selection rank
Sie besagt, dass der Kreuzungspunkt der Linien, die ein Bild vertikal und horizontal dritteln, der effektivste Platz für das Hauptmotiv ist.
Compare reference and output
Independent reference
Sie besagt, dass der Kreuzungspunkt der Linien, die ein Bild vertikal und horizontal dritteln, der effektivste Platz für das Hauptmotiv ist (siehe Beispiel).
Unchanged automatic output
Sie besagt, dass der Kreuzungspunkt der Linien, die ein Bild vertikal und horizontal dritteln, der effektivste Platz für das Hauptmotiv ist. Siehe Beispiel.
Weaker comparison case under the fixed selection rule
los undarbans han sido desde un templo de la vida silvestre y se calcula que en la actualidad existen en el área cuatrocientos tigres reales de bengala y unos treinta mil ciervos moteados
Compare reference and output
Independent reference
Los Sundarbans han sido, desde 1966, un templo de la vida silvestre y se calcula que en la actualidad existen en el área 400 tigres reales de Bengala y unos 30 000 ciervos moteados.
Unchanged automatic output
los undarbans han sido desde un templo de la vida silvestre y se calcula que en la actualidad existen en el área cuatrocientos tigres reales de bengala y unos treinta mil ciervos moteados ¡Gracias por ver el video!
Le rapport est très critique envers presque tous les aspects de la politique actuelle du gouvernement à l'égard de l'Irak et plaide vivement pour un changement de direction immédiat.
Compare reference and output
Independent reference
Le rapport est très critique envers presque tous les aspects de la politique actuelle du gouvernement à l'égard de l'Irak et plaide vivement pour un changement de direction immédiat.
Unchanged automatic output
Le rapport est très critique envers presque tous les aspects de la politique actuelle du gouvernement à l'égard de l'Irak et plaide vivement pour un changement de direction immédiat.
Weaker comparison case under the fixed selection rule
in questo porto si verificò lo stallo navale tristemente noto del in cui sette navi provenienti da germania stati uniti e gran bretagna si rifiutarono di abbandonare il porto
Compare reference and output
Independent reference
In questo porto si verificò lo stallo navale tristemente noto del 1889, in cui sette navi provenienti da Germania, Stati Uniti e Gran Bretagna si rifiutarono di abbandonare il porto.
Unchanged automatic output
in questo porto si verificò lo stallo navale tristemente noto del in cui sette navi provenienti da germania stati uniti e gran bretagna si rifiutarono di abbandonare il porto Grazie per la visione!
Hala hayatta olan buradan sağ çıkmayı başarabilmiş pek çok erkek ve kadın var.
Compare reference and output
Independent reference
Hala hayatta olan, buradan sağ çıkmayı başarabilmiş pek çok erkek ve kadın var. Çok sayıda kişinin sevdikleri, hem Yahudiler hem de Yahudi olmayanlar, burada katledildi veya ölümüne çalıştırıldı.
Unchanged automatic output
Hala hayatta olan buradan sağ çıkmayı başarabilmiş pek çok erkek ve kadın var. Çok sayıda kişinin sevdikleri hem Yahudiler hem de Yahudi olmayanlar burada katledildi veya ölümüne çalıştırıldı.
Weaker comparison case under the fixed selection rule
Uzak adalarda kredi kartları muhtemelen kabul edilmeyecektir.
Compare reference and output
Independent reference
Uzak adalarda kredi kartları muhtemelen kabul edilmeyecektir ancak Sterlin ve ABD Doları kullanılıyor olabilir. Kabul edilen ödeme yöntemini belirlemek için önceden mekan sahipleriyle görüşün.
Unchanged automatic output
Uzak adalarda kredi kartları muhtemelen kabul edilmeyecektir. Ancak sterlin ve ABD doları kullanılıyor olabilir. Kabul edilen ödeme yöntemini belirlemek için önceden mekan sahipleriyle görüşün. Çeviri ve Altyazı M.K.
Historical silence case from 29 September 2026: the model produced no text. Product code then added the placeholder shown here unchanged. This bug was fixed on 30 September 2026: a valid empty recognition result no longer creates an artificial cue. The old SRT documents the historical version; silence has no error rate.
Sie besagt, dass der Kräufungspunkt der Linien, die ein Bild vertikal und horizontal dreht, der effektivste Platz für das Hauptmotiv ist.
Compare reference and output
Independent reference
Sie besagt, dass der Kreuzungspunkt der Linien, die ein Bild vertikal und horizontal dritteln, der effektivste Platz für das Hauptmotiv ist (siehe Beispiel).
Unchanged automatic output
Sie besagt, dass der Kräufungspunkt der Linien, die ein Bild vertikal und horizontal dreht, der effektivste Platz für das Hauptmotiv ist. Siehe Beispiel.
Listen to the recording and compare the independent reference with the unchanged automatic SRT. All examples are accessible without an account.
Displayed examples: selection rank 1 and the highest character error rate in each audio language (lowest rank breaks ties; rank 2 if duplicated). The four additional diagnostics are separate from the 144 originals. Both good and weaker results remain visible.
Short read speech. No general accuracy guarantee, human listening approval or verified synchronization. Long conversations and overlapping voices are not comprehensively tested. This small test establishes neither superiority, certification nor broad everyday suitability. Overlap with model training data cannot be ruled out; reliable speaker IDs are unavailable.
144 short original recordings, 28.89 minutes in total: 12 of 12 outputs per language, without technical failure. All used the complete processing code in isolation; only one Japanese case also used regular upload and SRT download. No human listening or timing approval.
Selected before recognition: official FLEURS test split, seed 20260929, fixed SHA-256 ordering; twelve distinct recordings and reference sentences per language. References were not supplied to recognition. Scoring uses Unicode NFC, lowercase and normalized punctuation; numbers, negations, diacritics and writing systems are not semantically rewritten. Number words versus digits and spacing can increase error rates without changing meaning. No ranking across languages.
Standard quality and automatic language detection: large-v3, Beam/Best-of 3, Standalone Faster-Whisper-XXL r245.2 / Faster-Whisper 1.1.1. Preprocessing and postprocessing, including content-dependent CJK checks, remained active. No translation, optional audio enhancement or speaker separation.
WER: word error rate. CER: character error rate. Substitutions + deletions + insertions divided by reference size. JA/ZH: CER only. Rates cover successful outputs; errors: technical / missing output. No accuracy as 100 minus WER.
CSV: all 148 cases labeled separately. JSON: denominators, error operations and distribution (minimum/median/maximum). CER ignores spaces; Korean WER uses whitespace units. Original texts and numbers remain traceable.
Historical silence case from 29 September 2026: the model produced no text. Product code then added the placeholder shown here unchanged. This bug was fixed on 30 September 2026: a valid empty recognition result no longer creates an artificial cue. The old SRT documents the historical version; silence has no error rate.
The additional closing phrases remain unchanged pilot results. Detection of these phrases is still an unpublished prototype; no measured improvement is claimed here.
Results for all twelve audio languages and error rates
144 short original recordings, 28.89 minutes in total: 12 of 12 outputs per language, without technical failure. All used the complete processing code in isolation; only one Japanese case also used regular upload and SRT download. No human listening or timing approval.
Audio language
Evaluated / selected
Audio minutes
WER
CER
Errors
Deutsch · de_de
12 / 12
2.41
4.37 %
1.74 %
0 / 0
English · en_us
12 / 12
1.93
3.15 %
1.59 %
0 / 0
Español · es_419
12 / 12
2.16
4.74 %
4.22 %
0 / 0
Français · fr_fr
12 / 12
2.06
6.46 %
2.48 %
0 / 0
Português · pt_br
12 / 12
2.43
2.34 %
0.34 %
0 / 0
Italiano · it_it
12 / 12
2.83
3.10 %
1.67 %
0 / 0
日本語 · ja_jp
12 / 12
2.82
—
3.23 %
0 / 0
한국어 · ko_kr
12 / 12
2.98
7.80 %
0.79 %
0 / 0
Türkçe · tr_tr
12 / 12
2.62
5.12 %
1.81 %
0 / 0
Русский · ru_ru
12 / 12
2.55
5.88 %
1.51 %
0 / 0
العربية · ar_eg
12 / 12
2.11
12.26 %
3.28 %
0 / 0
中文 · cmn_hans_cn
12 / 12
1.99
—
3.24 %
0 / 0
WER: word error rate. CER: character error rate. Substitutions + deletions + insertions divided by reference size. JA/ZH: CER only. Rates cover successful outputs; errors: technical / missing output. No accuracy as 100 minus WER.
German video demo and additional translation
From speech to subtitles
Spoken language: German
Subvideo.ai
From speech to subtitles
A German recording was supplied as an MP4 with a plain background and transcribed by Subvideo.
Transcription first produces German subtitles. Translation is a separate product step; a normal upload does not automatically translate into the page language.
Prepared presentation, not a Studio recording or product-styled video export.
Preview translation: editorially reviewed and edited with AI assistance. The unchanged automatic result is available separately.
Unedited automatic translations may contain errors or omissions. Text and timing need review.