Parallel Recordings

This dataset is a small corpus of audio recorded in parallel. This data can be used to quantify the difference in recording quality between different audio devices. An example of this quantification using WER is shown below.

Organization

There are two separate experiments.

Personal Microphone
Speakerphone

The original data for both experiments are located in thefull_length_audio directory. Each experiment was performed by recording simultaneously on multiple devices. Example sentences were prepared beforehand and read during the simultaneous recording. The recordings were cut into bits that could fit within the context of most modern ASR models (ie. < 30s).

Personal Microphone

Devices:

AT8: Audio Technica Digital 2.4GHz System 10 + ATW-T1007 + AT857SL
ATR: Audio Technica ATR4800-USB
bltin: Built-In microphone on unknown laptop
boya: Boya BY-W4, Ultracompact 2.4GHz Four-channel Wireless Microphone System

Each device was pointed in the same direction toward the speaker. The speaker read sentences at varying distance from the microphones. Here is the following mapping:

Indexes 4-13: Distance = 1 inch (2.54cm)
Indexes 18-27: Distance = 1 foot (30.48cm)
Indexes 30-39: Distance = 1 yard (914.4cm) All other indexes correspond to user speaking to themselves at varying and inconsistent distances.

Example of Usage

This data was used to measure microphone quality in the context of ASR. The metric used was WER from the output of WhisperModel from faster_whisper. The follow parameters were used:

whisper = WhisperModel(
    model_size_or_path="deepdml/faster-whisper-large-v3-turbo-ct2",
    device="cuda",
    compute_type="float16",
)

Speakerphones

Devices:

respeaker: SeeedStudio ReSpeaker USB Mic Array
jabra: Jabra Speak2 75, PHS060Wa
sanwa: Sanwa Supply 400-MC011
yamaha: Yamaha Unified Communications Speakerphone YVC-330
ntt: NTT Advanced Technology FR-1100

Each device was placed in the same location in the center of the room. The speaker read sentences five at a time, moving position each time. The positions were roughly assigned and can be best described as close-rear-left, far-front-left, far-front-right, close-rear-right (in order). The final paragraph was read in the same location as the last 5 sentences (close-rear-right).

The NTT microphone has directionality settings and was set to capture as wide as possible but there are certain angles where complete loss of sound can be observed. This was also recorded originally as a stereo track.

audio audioduration (s) 2.41 277