Beth Worthy
7/6/2021
Before we understand how to customize your data collection project, let's first discuss what speech collection really is.
Speech data collection is a process of collecting and measuring high-quality audio data from different sources.
Data collection is crucial for feeding the Automatic Speech Recognition (ASR) system. The system is responsible for understanding and translating different languages within seconds.
However, for the system to accurately interpret and translate the audio data, it must first be exposed to a high volume of high-quality speech data.
Before you start collecting and measuring voice data, you need to make some choices about your project.
While some clients know exactly how they want to structure their voice data, some might come with loose requirements either because they are unaware of all the possible variables or have flexible requirements.
This blog deals with tailoring speech collections based on languages, demographics, collection size, script structure, audio quality, etc.
These customizations will ultimately affect the data collection method, number of participants required, delivery timeline, and project cost.
Let's now look into how speech data can be efficiently collected and customized for end users.
The first step towards speech data customization is to clarify what language needs to be collected.
You must know whether the participants are expected to be native, non-native, or a mixture of both.
Knowing this before the project commences would speed up the entire operation while helping to improve the quality of collected data.
You must also know whether you want to collect the speech notes in a different or foreign accent. You can customize your collection to record different dialects from participants across diverse regions.
For instance, to avoid exposing the speech recognition algorithm to systematic bias, you can include participants who speak in a variety of dialects in different accents.
You must also assess whether you want to collect voice notes from a specific country. If not, you can include accents from several countries where the selected language is not official.
For instance, the accents of Spanish speakers from Mexico and Spain are very different.
So, you should determine whether you want to expose your algorithm to different language accents or stick to a specific region.
You can also customize your data collection with demographic variables. Start by targeting specific male vs. female distribution or children vs. adults.
These would make up your target demographic for speech collection and measurement.
Generally, these language and demographic variables can be summed up by a brief declaration about your target demographic, e.g., collecting voice notes worldwide from native and non-native Spanish speakers
You can also customize your script structure based on your collection needs. You can choose to create one unique script for all participants or different speeches for different participants.
For instance, you can make one section of female participants read one script while making the others read a different one.
Scripted speech would direct participants to read aloud what they see on a screen. Natural language would provide the participant with a specific scenario, and they can narrate their thoughts based on it.
You must also be clear on the total utterances you need for your algorithm to understand and store data.
The higher the data requirement, the greater the need for the number of participants and their utterances would be.
For example, 25 participants with 60 utterances each would mean 1500 total repetitions. Depending on the type of speech data, you can customize the participant's speech script.
The size of participants would differ depending on the amount of data you want to feed your ARS system. It would also impact the number of words per participant and the number of participants required.
Based on collection size, you can understand your participant needs. If you are collecting several languages, then evaluate the number of participants required per target language.
And if you are collecting data based on demographics, set your breakdown per demographic accordingly.
Watch out for distracting background noises while recording voice notes. It can impact the quality of collected data and consequently intervene in its accurate interpretation by a voice recognition algorithm.
You might have certain structure or post-processing requirements for the project’s audio recordings or files. This could mean needing leading or trailing silences or deletion of noises like taps or clicks.
Moreover, you might require certain files to be combined together. You will need to define these requirements before the survey process commences clearly. You can also customize the way the audio files are sent to you.
If you need the speech data to be transcribed or labeled before delivery in accordance with a specific set of labeling, segmentation, or noise-marking guidelines, hire a professional translation services provider.
Additionally, if you face difficulty preparing a speech transcript in a different language or are unable to transcribe the relevant data for your target language on your own, you can always opt for transcription services.
It would enable you to get the content professionally transcribed in your desired format and target language for your speech data collection process.
Do you need to collect data for a speech collection project? Hiring professional transcription services like GMR Transcription Services, Inc. can boost your speech collection and customization technique.
Our pool of experienced and highly skilled transcriptionists can handle small-scale and large-scale audio transcription projects with 99% accurate transcripts while adhering to quick deadlines.
Contact us about our research audio and document transcription services!
Also Read: Speech Recognition and Mobiles
