The whole pipeline will automatically generate abundant types of tags and transcriptions on the collected audio data. The human-in-the-loop scheme enroll native language speaker as human annotator to audit all the automatical tags and transcriptions and correct if necessary. In our SaaS platform client.appen.com, we set up audit jobs for native language spoken workers to do review and finalize the annotation. The instructions provide to them is show in below:
OTS Automation- Annotation Job
Overview
Listen to a short piece of audio and transcribe the text.
We Provide
- A short piece of audio
- A text box to enter the transcription Process
- Step 1: Listen to the audio
- Step 2: Verify audio tags
- Choose the noise type of the speech in the audio
- Determine whether the speech of entire audio is in US English
- Determine whether there is only one speaker or more in the audio
- Determine the gender of speaker if there is only one speaker
- Determine whether the content in this Audio is offensive
- Step 3: Review and Correct Audio Transcription Review the transcription in text box, correct it if necessary.
Notes on Transcriptions
This batch contains a hypothesis in the text box. Edit the transcription to match the audio.
- Please pay particular attention to correcting the use of similar-sounding words: hi/high, there/their, etc.
| Category | Audio Clip | Correct Transcription |
|---|---|---|
| Self-correction• When the user self-corrects, transcribe everything the user says | Will would you open spotify | will would you open spotify |
| Numerals• Words or phrases should be the spelled-out version | 891 oh 2 | eighty nine one o two |
| Time and Date Phrases• Words or phrases should be transcribed as the user says | 7:00 1/2/2019 | seven o’clock january second twenty nineteen (or january second two thousand nineteen based on the audio |
| Contractions• When encountering contractions, do not split the contraction into two words. We should be transcribeing exactly what we hear | i'm gonna leave i'm going to leave | i'm gonna leave i'm going to leave |
| Acronyms• Acronyms that are pronounced as an actual word instead of as a sequence of letters will be transcribed as one word | NASA SCUBA | nasa scuba |
Pronunciation
- Speakers often pronounce his as “he’s”. This is just the accent, so it should still be transcribed as his
- People will often say they gone when speakers of other dialects might say “they’ve gone” or “they left”. Transcribe the words exactly as they are spoken
- we'll and will may sound closer to each other than in other accents you’re used to. Make sure you are using the correct one
- that’s may sound more like “thass” or “dass”. Be sure to still transcribe it as that’s
- address can sound a bit like “edges”. Be sure to spell it address in that case
- phone can sound like “fuh-wan” or “fowan”. Be sure not to accidentally transcribe it as a word like “following”
Grammar
- Be mindful of the difference between your (possession) and you’re (you are)
- Be mindful of the difference between its (possession) and it’s (contraction of “it is”) and be sure you are using the right one
- Please spell okay rather than “ok”
- Keep in mind the difference between their (possession) and they’re (contraction of “they are”) and make sure you are using the right one
- Bear in mind the difference between suites and sweets, which are pronounced the same. If the person is talking about a hotel or other location, the correct spelling will most likely be suites
- If a speaker says something like “buh-bye” please transcribe it as bye bye