An audio CAPTCHA either stands alone ("type what you hear") or serves as the accessibility alternative to a visual one such as reCAPTCHA, hCaptcha, AWS WAF or MTCaptcha. The clip typically reads out digits or letters over background noise, music or distortion. For automation, this fallback is often the easiest way around an otherwise tough visual challenge.
How to identifyLink to How to identify
- A speaker or audio-player icon, frequently next to a visual challenge
- Instructions such as "Type what you hear" or "Press play"
- Requests to paths like
/audio,/audio.mp3or/captcha/audio, or base64 audio inside JSON - Responses typed
audio/mpeg,audio/wavoraudio/ogg - A refresh button, because the first clip may be made deliberately noisy
How it worksLink to How it works
- The server picks a random answer and turns it into a short clip, usually text-to-speech plus noise.
- The clip's URL belongs to the session and expires fast.
- The visitor types the answer and the server compares it with the stored value.
- Some systems change the answer after a number of failures, or cut off audio after too many requests.
Free bypass methodsLink to Free bypass methods
| Tool | What it does | Success Rate | Link |
|---|---|---|---|
| Whisper | The best general option; large-v3 copes well with noise | 90-99% | openai/whisper |
| faster-whisper | A CTranslate2 port, about 4× faster and lighter on memory | 90-98% | SYSTRAN/faster-whisper |
| Vosk | Small offline speech-to-text, good with digits and letters | 70-85% | alphacep/vosk-api |
| wav2vec2 | Self-supervised speech models on Hugging Face | 70-90% | facebook/wav2vec2-base-960h |
| SpeechBrain | Speech toolkit with pretrained models | 70-90% | speechbrain/speechbrain |
| noisereduce | Spectral-gating noise removal ahead of transcription | — (preprocessing) | timsainb/noisereduce |
| FFmpeg | Converts, trims, normalises and resamples audio | — (preprocessing) | FFmpeg/FFmpeg |
Recommended pipelineLink to Recommended pipeline
- Download the clip using the same session cookies as the challenge.
- Convert it to 16 kHz mono WAV:
ffmpeg -i in.mp3 -ar 16000 -ac 1 out.wav. - Cut the silences, even out the volume with
loudnorm, and denoise lightly if needed. - Transcribe with faster-whisper, choosing
small,mediumorlarge-v3according to how noisy the clip is. - When the site only uses digits, filter the transcript down to that alphabet.
- If the answer is refused, ask for a new clip. Retrying the same one rarely works.
Improving accuracyLink to Improving accuracy
- Run a few versions of the clip (raw, denoised, high-pass filtered) and keep the most confident transcript.
- Set a language hint such as
language="en"and turn translation off. - For digit-only clips, turn spoken numbers ("seven") into digits and remove spaces.
- If the speech is slow, split the clip, transcribe each part and join the results.
- Keep a small local model for volume work and a large one for the hard clips.
Practical tipsLink to Practical tips
- Audio endpoints get rate-limited and flagged sooner than visual ones. Spread requests across IPs.
- reCAPTCHA's audio route is heavily abused, so Google may serve garbled clips or shut audio off for suspicious sessions.
- Some sites only enable the audio button while the page is focused and visible, so behave like a real visitor.
- Answers are usually case-insensitive, but check how the site behaves before normalising.
- Solve and submit within the same session, and be quick: audio tokens expire fast.