AI Audio Transcriber — Free Speech-to-Text Online

MP3, WAV, M4A, AAC, FLAC, OGG and Opus are accepted directly. Other formats can be normalised first with Audio Converter.

No. The audio is decoded locally and fed to whisper.cpp running inside your browser tab. The transcript is generated in memory and downloaded to your device only.

Whisper-tiny reaches 90%+ accuracy on clean English audio. Heavy accents, background noise, overlapping speakers and music all reduce accuracy. Run the file through Audio Noise Reducer first if the recording is noisy.

Yes — toggle the "Include timestamps" option to export segment-level start and end times alongside each line.

Up to 200 MB. A 60-minute MP3 at 128 kbps fits comfortably under that ceiling.

Whisper does not know domain-specific vocabulary that was rare in its training data — niche pharmaceutical names, internal product codenames, obscure programming terms, etc. It will usually substitute the closest common word, which is easy to fix in the preview editor before you export. For repeated jargon you can do a Find & Replace on the exported .txt afterwards.

No. Whisper.cpp is a transcription engine, not a speaker-diarization engine. Every line is recognised but the transcript does not say who said which line. If you need speaker separation, you can manually annotate during the editing pass — that is still substantially faster than typing the entire recording from scratch.

No. The file is read into your browser tab, decoded by the WebAudio API, and processed by whisper.cpp running locally. The transcript is generated in memory and offered as a direct download. Closing the tab clears every byte; nothing is logged on any server.

Roughly 0.2-0.4× the audio duration on a modern laptop, so a 30-minute recording finishes in 6-12 minutes. On phones the ratio climbs to 0.4-0.8× — expect 12-24 minutes for the same 30-minute clip. The first run also includes the one-time engine + model download (~39 MB combined).

Aim for 16 kHz or higher sample rate, mono, and at least 64 kbps if you must compress. Lossless formats like WAV and FLAC give the best accuracy because they have no MP3-style codec artefacts on consonants. If you are recording on a phone, the default Voice Memos format on iOS (M4A) and the default Recorder format on Android both work well.

One at a time. The model holds the audio buffer plus the inference state in memory; running two simultaneously would likely OOM the tab on most devices. Drop the next file in once the first transcript is downloaded — the model is already loaded so each subsequent file skips the engine warm-up step.

The pipeline is functionally similar — both use a deep-learning speech model — but cloud services run inference on the provider’s servers, with their own pricing and retention policies. The AI Audio Transcriber runs the same family of model entirely on your device, with no per-minute charge and the audio never leaving the browser. The trade-off is that some cloud services use larger, slower-but-more-accurate models on dedicated GPUs.