What Is the Whisper Model?
Whisper is an automatic speech recognition system released by OpenAI. It is designed to transcribe spoken language into text. The model was trained on a massive dataset of diverse, weakly supervised audio, giving it robustness against background noise, varied accents, and technical jargon.
Unlike proprietary cloud-only APIs, OpenAI open-sourced the Whisper model weights and code under the MIT License. This open-source distribution allows developers to run the inference engine independently without sending audio data to third-party servers.
How Whisper Processes Audio
Whisper employs an encoder-decoder Transformer architecture. When an audio file is provided, it is first converted into a log-Mel spectrogram, which is a visual representation of the audio frequencies over time.
The Transformer encoder processes this spectrogram to understand the acoustic features, while the decoder generates the corresponding text transcript. During this process, the model can also predict timestamps, allowing the text to be aligned with the precise moment it was spoken.
Whisper Implementation in Konthora
Konthora utilizes the Whisper model to power its free, browser-based transcription tool. Specifically, Konthora implements the small.en variant of the Whisper model, which provides an optimal balance between transcription precision and computational speed for English-language audio.
The Konthora implementation introduces specific guardrails and processing features to support practical user workflows:
Supported Formats
- Audio: MP3, WAV, M4A, AAC
- Video: MP4, WebM, MOV
- Export: TXT, SRT, VTT, JSON
Processing Constraints
- 10-minute maximum duration
- 100 MB maximum file size
- English language only
Privacy and Data Lifecycle
Because the Whisper model is open-source, Konthora hosts the inference engine independently. Audio files are not transmitted to OpenAI for processing.
Within the Konthora infrastructure, all uploaded media and generated transcripts are subject to a strict 60-minute automatic deletion policy. The tool does not require user accounts, ensuring that transcriptions remain untethered to persistent user identities.
Using Whisper for Captions
One of the most practical applications of the Whisper model is the generation of timestamped subtitles and closed captions. Through the Konthora interface, the raw timestamps produced by Whisper can be grouped by word, sentence, or paragraph, and exported directly as SRT or VTT files.