Speech to Text Guide

Transcribe voice notes, meetings, and audio files to clean text fast.

How to use

Tips

Model and network behavior

The tool uses a multilingual Whisper tiny int8 model through sherpa-onnx WebAssembly. The first run downloads the runtime and more than 100 MB of fixed model files through Aiviko's same-origin model route. That route retrieves only the pinned model assets; it does not receive the audio you choose. Later visits may reuse the browser cache.

The language menu reflects languages understood by the upstream model, not a promise of equal accuracy. Tiny models trade accuracy for download size and local speed. Accents, overlapping speakers, music, specialist terms, code-switching, distant microphones, and compressed recordings can reduce quality. Automatic language detection may be wrong on short clips; choose the spoken language when known.

Review and troubleshooting

Treat the transcript as a draft. Check names, numbers, quotations, medical or legal terms, and speaker changes against the recording. If initialization fails, verify that the browser allows WebAssembly and workers and that network filtering is not blocking model requests. Close other memory-heavy tabs for long audio. A faster desktop can finish work that a phone cannot process reliably.

Privacy

Speech to Text decodes and transcribes audio in your browser. The selected audio is not uploaded to an Aiviko processing server; runtime and model assets are separate network downloads.

Limits and troubleshooting

WAV, MP3 and M4A are common inputs; successful decoding depends on the codec and browser. The interface language is separate from the spoken-language setting. Choose the spoken language for short clips instead of relying on automatic detection. For timed SRT or VTT output, use the subtitle generator.

Related tools

• Subtitle Generator
• Text to Speech