Voice
Two models pointed in opposite directions. Synthesise a voice with adjustable rate and pitch, then feed real audio to a transcriber and see where it mishears.
Press speak. The waveform above is decoded from the actual audio — the tall sections are vowels, the narrow spikes are consonants, and the flat gaps are the pauses the model chose to insert.
Nothing transcribed yet
Record a sentence or upload a file. Try a proper noun or a technical term and watch it substitute something more common.
Whisper tiny is 39 million parameters — the smallest of six sizes, chosen here because it runs on a modest server. The larger versions are markedly better on accents, names and technical vocabulary.
Transcription converts sound into a spectrogram — a picture of which frequencies are loud over time — then treats reading that picture as a translation problem, emitting text tokens. Synthesis runs the pipeline in reverse.
Recognition accuracy varies substantially across accents and dialects, reflecting what was in the training audio.
- Whisper tiny is the smallest of six sizes. It trades accuracy for speed, and it shows on accents, names and technical vocabulary.
- It never returns an empty transcript. Given silence or noise it will produce plausible words anyway.
- Synthesised prosody comes from punctuation. Move a comma and the phrasing of the sentence changes.
This page says “the model predicts”, not “the model knows”. That is deliberate. None of these systems understand the images or sentences you give them; they map inputs to outputs using patterns fixed at training time. The difference matters most exactly when the output is impressive.