Sample Audio Files for Transcription Testing
Download Sample File
Free test file, safe content, instant download.
7e9bef643ddb97e34ba41df7...Testing a speech-to-text pipeline needs actual speech — tones and music do not exercise a model the way a voice does. This page ships three purpose-built speech files plus their transcripts: a 5-second clean WAV, a 38-second long-form WAV for sustained input, and the same long recording degraded with band-limiting and pink noise for hard-mode robustness testing.
The clean short WAV is deliberately in the exact format most ASR systems natively consume: 16 kHz sample rate, 16-bit PCM, single channel — the same input Whisper uses internally. Because the ground-truth transcript is documented (sample-speech-transcript.txt), you can compute exact word error rate (WER) on a known reference instead of eyeballing output.
Every file was synthesized by us (eSpeak NG formant synthesis + FFmpeg processing) and released under CC0 — no third-party recordings, no human voice, no consent or privacy questions. Ship them in test fixtures, CI pipelines, and demos without attribution obligations.
File Information
- File 1
- sample-speech-short.wav — 5.1s, 16kHz mono PCM, clean synthetic speech
- File 2
- sample-speech-long.wav — 38.5s, 16kHz mono PCM, sustained input
- File 3
- sample-speech-noisy.wav — same 38.5s, band-limited + pink noise (hard mode)
- Transcripts
- sample-speech-transcript.txt — exact ground truth for WER
- License
- CC0 — synthesized in-house, zero third-party recordings
- Best for
- Whisper, Deepgram, AssemblyAI, Google STT, WER evaluation
How to Use This Audio (Speech) Test File
- 1Download sample-speech-short.wav and sample-speech-transcript.txt (the known-transcript reference).
- 2Run it through your ASR pipeline or Whisper CLI: whisper sample-speech-short.wav --model tiny.
- 3Compare output against the transcript to measure word error rate.
- 4Re-run with sample-speech-long.wav (sustained input) and sample-speech-noisy.wav (noise robustness).
Frequently Asked Questions
- What makes a good audio file for transcription testing?
- Actual speech (not tones or music), a documented transcript for WER measurement, 16 kHz or higher sample rate, and ideally both a clean and a noisy variant. Our files cover short, long, and noisy cases — and because everything is released CC0, you can embed them in test suites without license review.
- Why a synthetic voice instead of a human recording?
- Rights and reproducibility. A human recording carries copyright, consent, and (for real-world captures) privacy questions that make it risky to redistribute; a synthetic voice has none of that. The trade-off — a slightly robotic timbre — is a feature for testing: it is a consistently hard input that stresses a model’s acoustic front-end the way clean studio speech never will.
- What is the transcript of the short WAV file?
- “This sample voice recording is used for testing speech recognition systems.” The full ground truth for all three files is in sample-speech-transcript.txt — documented by construction, since we generated the audio from the text. WER computation is trivial: one reference, no ambiguity.
- Can I use these files with the Whisper API?
- Yes. The WAV is 16kHz mono PCM — Whisper’s native format. Via the OpenAI API: POST it as a file to /v1/audio/transcriptions with model whisper-1. Via the local CLI: whisper sample-speech-short.wav --model base. All files also work with Deepgram, AssemblyAI, and Google Speech-to-Text.
- Are these speech files free to use in commercial test suites?
- Yes — CC0 1.0 Universal, no attribution required. The audio was synthesized locally (eSpeak NG formant synthesis) from text we wrote; there are no third-party recordings or licensed voices in the chain.
Related Sample Files
Need a different size or format?
Use our free Generator to create custom test files in any format and size — right in your browser, no upload needed.
Try Our Generator