OpenAI: Whisper Large V3

Whisper large-v3 is a released OpenAI speech-recognition checkpoint with approximately 1.55 billion parameters. It is a multilingual encoder-decoder model: audio enters the encoder and the decoder predicts text and task-related tokens. The principal tasks are speech transcription in the source language and translation of speech into English. It is an open-weight model identity, kept separate from a hosted service's model identifier or that service's upload and response-format rules.

Large-v3 retains the architecture of earlier large Whisper checkpoints with two stated changes: its spectrogram input uses 128 Mel-frequency bins instead of 80, and it adds a Cantonese language token. The producer's model card describes training on one million hours of weakly labelled audio and four million hours of pseudo-labelled audio prepared with Whisper large-v2. This mixture was used for two training epochs. The release is a specific large-v3 checkpoint rather than a generic name for every model in the Whisper family.

The documented inference workflow performs transcription by default, preserving the language spoken in the recording. It can detect the source language automatically, or accept a known language selection. For translation, the multilingual large model provides an English text result from non-English speech. Choose that task explicitly: a same-language transcript and an English translation are different artifacts. For an interview, retain the original-language wording; for an English reference document, specify that the desired result is a translation of the spoken content.

Audio preparation is part of the input path. OpenAI's loader decodes a recording to a mono waveform and resamples it to 16 kHz. The model's acoustic receptive field is 30 seconds. That is a processing window, not a maximum length for an entire interview or lecture. Long-form inference can advance sequentially through sliding windows, or split a long recording into overlapping chunks and join their transcripts. The producer's model card describes sequential processing when accuracy is the priority and chunked processing when the turnaround of a single long recording matters more.

The released configuration lists 448 decoder positions for an individual decoding sequence. Longer recordings are handled by the long-form process rather than by treating that number as a universal transcript limit. The documented workflow can return segment-level or word-level timestamps alongside recognized text. Those timing outputs are useful for navigating recordings and preparing captions, while any final subtitle container or transcript-file format is a separate output-formatting choice. Preserve the recording's timing when preparing audio if the transcript will be aligned with the original material.

Whisper's intended role is recognizing and translating speech. A lecture transcript, an interview record or a searchable collection of spoken material is a practical output; summarization or interpretation can follow as another step over the text. The model card does not establish robust automatic speaker classification or diarization as a standard output of the base checkpoint. If a task needs speaker labels, that requirement must be handled separately from claiming an accurate transcript. A useful brief identifies the source language, whether transcription or English translation is needed, and whether timing information is required.

Substantive limitations include invented text that was never spoken and repetitive text, associated with weakly supervised training on noisy data and the sequence-to-sequence design. Recognition quality varies across languages, accents and dialects, with weaker results where less training material exists. Noise, specialist vocabulary and overlapping speech should therefore be represented in the sample used to judge suitability. Compare names, quantities and important statements with the recording, and inspect non-speech passages or repeated lines rather than treating a fluent transcript as proof that every word was heard.

A complete result retains recognized text, its language and any requested timing with a reference to the original recording. Keep transcription separate from later editing: a corrected publication transcript may remove fillers, while the recognized transcript preserves what the model actually returned. Large-v3's identity and translation capability should also remain separate from the faster Turbo edition. Actual processing speed depends on the language, speaking rate, hardware and inference method; no particular hosted latency, maximum upload size or enabled application route follows from the checkpoint name alone.