Skip to main content
Use this recording for the SDK quickstarts. The noisy and clean versions contain the same 56,080 samples: mono PCM16 WAV at 16 kHz, lasting 3.505 s. Each file is about 112 kB. Or download the noisy file from a terminal:
The recording says: “Concord returned to its place amidst the tents.”

​
Check the input

Verify the checksums and keep both source files separate from the enhanced output.

​
Attribution and changes

The source is LibriSpeech test-clean utterance 6930-75918-0000, from LibriSpeech: An ASR Corpus Based on Public Domain Audio Books by Vassil Panayotov, Guoguo Chen, Daniel Povey and Sanjeev Khudanpur. OpenSLR SLR12 distributes the corpus under Creative Commons Attribution 4.0 International. These derived fixtures retain that license. No endorsement by the source authors is implied. Source archive: test-clean.tar.gz; member: LibriSpeech/test-clean/6930/75918/6930-75918-0000.flac. The source FLAC SHA-256 is 9ce35224156f071ab58eb7feb8a5ceae600f6f9f353da2a6cbf797b6b1ac8a23. Changes made for this example: decoded the FLAC to mono samples, exported PCM16 WAV and created a second version with seeded Gaussian white noise. Noise was scaled to a 12 dB signal-to-noise ratio over the whole clean recording before PCM quantization, using NumPy 2.5.3 default_rng(20260918) and SoundFile 0.13.1. No speech was cut, reordered or transcribed automatically.

​
Use representative audio for evaluation

This short recording with synthetic noise checks setup. Evaluate quality on recordings that represent your microphones, languages, codecs and competing speakers. After setup, follow audio evaluation with recordings representative of your application and review audio format before changing rates or channels.