BACKAryan Sakhala
Free speech to text
Speak, or drop a file. Get the words.
Record from your microphone with a live transcript, or upload audio and video files up to 60 minutes. Qwen3-ASR runs on Qwen's free Hugging Face Space. Nemotron 3.5 runs on your own GPU inside the browser, so your audio never leaves your device.
How does speech recognition work? See the architectureAudio
—
Wall time
—
Speed
—
Language
—
Transcript will appear here. Switch models and re-run the same file to compare them.
How does Whisper architecture work?
Under the hood
How speech becomes text.
Modern speech models, such as Whisper and the models on this page, use an encoder that listens and a decoder that writes. The decoder keeps two caches, so it does not repeat work for every new word.
- Audio Your recording becomes 16,000 numbers per second.
- Spectrogram The sound becomes a time-frequency heat map.
- Encoder The model turns the heat map into audio features, once per chunk.
- Decoder The model writes one token at a time. It reads the audio through cross-attention.
- Text The tokens become words, punctuation and timestamps.