Whisper is a general-purpose speech recognition modelCITE:E4. It is trained on a large dataset of diverse audio and is also a multitasking model that can perform multilingual speech recognition, speech translation, and language identificationCITE:E5. That combination — one model covering recognition, translation, and language identification rather than separate specialized systems — defines its stated scope.
OpenAI trains Whisper to predict transcripts of large amounts of internet audio, the core method behind its weak-supervision approachCITE:E1. Structurally, Whisper is built as an encoder-decoder (sequence-to-sequence) transformer, pretrained on 680,000 hours of labeled audio dataCITE:E6. In other words, the training signal comes from transcripts already present on the internet rather than from a purpose-built, manually annotated corpus, and that signal is fed into a standard encoder-decoder transformer architecture.
When scaled to 680,000 hours of multilingual and multitask supervision, Whisper generalizes well and is often competitive with fully supervised results in a zero-shot transfer setting, without any fine-tuningCITE:E2. This amount of pretraining data enables zero-shot performance on audio tasks in English and many other languagesCITE:E7. The same 680,000-hour figure appears consistently across the original research paper and separate technical documentation, describing both the scale of the training data and the resulting zero-shot capability.
Whisper's models approach human-level accuracy and robustness when directly compared against human transcribersCITE:E3. This comparison is presented alongside the zero-shot, no-fine-tuning results described above, tying the training method to the reported outcome.
Across the research paper, the model repository, and independent technical documentation, the same figures recur: transcripts of internet audio as the training signalCITE:E1, 680,000 hours of multilingual and multitask supervision as the scaleCITE:E2CITE:E6, and zero-shot performance across English and many other languages as the outcomeCITE:E7. Whisper's general-purpose, multitasking framingCITE:E4CITE:E5 and its near-human accuracy and robustness claimCITE:E3 are presented as the product of that same 680,000-hour weak-supervision process, not as separate achievements.
Whisper is a general-purpose speech recognition modelCITE:E4.
OpenAI trains Whisper to predict transcripts of large amounts of internet audio, the core method behind its weak-supervision approachCITE:E1.
When scaled to 680,000 hours of multilingual and multitask supervision, Whisper generalizes well and is often competitive with fully supervised results in a zer…
Whisper's models approach human-level accuracy and robustness when directly compared against human transcribersCITE:E3.
Whisper's core bet is architectural conservatism paired with data scale: the underlying model is a standard encoder-decoder transformer, and its stated advantage — zero-shot performance competitive with fully supervised systems — is attributed to the 680,000 hours of weakly supervised transcript data rather than to a novel architecture. That is a notable claim because it implies transcription quality and multilingual/multitask coverage track data volume more than model design. The detail worth watching is one the source material itself leaves open: the zero-shot claim is stated only as covering 'English and many other languages,' with no per-language breakdown or accuracy figure attached. Until that gap is filled, the 'near-human accuracy and robustness' comparison should be read as a benchmark-level claim rather than a guarantee across every language Whisper's documentation says it supports.
Google announced a €13 billion ($15.1 billion) investment in Finnish AI infrastructure to be deployed in 2027–2028, its largest single investment in Europe. The plan covers new data centers in Kajaani, Muhos, and Vaala plus an expansion in Hamina, backed by a 22-year power deal for 50% of Loviisa nuclear plant output and projected to support 37,000 construction-phase jobs and 7,000 permanent roles.
Chenbro (勤誠, 8210) reported August 2026 revenue of NT$2.457 billion, roughly flat month-on-month (-0.2%) but up 40.5% from a year earlier. Cumulative January–August revenue reached NT$19.848 billion, up 51.7% year-on-year.
Chang Hwa Bank (TWSE: 2801) reported cumulative after-tax profit of NT$15.567 billion for January through August, up 23.21% year-on-year with EPS of NT$1.29, a record for the period. The bank attributed the gain to loan, deposit, and wealth management momentum, and August alone delivered NT$2.916 billion in profit, up 29.16% year-on-year.