引導式測驗 · 中等 等級
語音與音訊資料標註測驗
查看轉錄約定、二值化、時間戳記、重疊以及基於參考的 WER 與更廣泛的註釋品質有何不同。
問題 1 的 8
When are word-level timestamps useful in a speech dataset?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- When are word-level timestamps useful in a speech dataset?
- When a downstream task needs timing for individual words; not every ASR corpus requires them
- Only when no transcript is available
- To identify a speaker’s real-world name
- To replace the transcription style guide
- What does speaker diarization labeling accomplish that plain transcription does not?
- It removes filler words automatically
- It identifies who is speaking during each segment, even without knowing their real identity
- It translates the audio into another language
- It measures the volume level of the recording
- Why do speech annotation projects use style guides for transcription?
- To make transcription illegal without one
- To replace the need for timestamps
- To standardize judgment calls, like how to handle filler words or numbers, across many annotators
- To automatically detect the speaker's identity
- How do annotation tools commonly help annotators place precise timing markers?
- By requiring annotators to guess based on hearing alone
- By automatically deleting all timestamps after labeling
- By translating the audio to text with no human involvement
- By showing a waveform or spectrogram alongside the audio so markers can be placed visually
- How should annotators handle overlapping speech?
- Apply the project’s specified overlap convention and preserve uncertainty
- Always omit one speaker without recording it
- Guess which speaker is more important
- Merge both turns into one word sequence without speaker labels
- What does Word Error Rate compare?
- Two speakers’ identities
- A hypothesis transcript against a designated reference transcript
- Whether every timestamp is medically accurate
- A clip’s audio quality against a transcript length
- Why is transcription described as more than purely mechanical dictation?
- Because it requires no audio at all
- Because transcription always removes all speaker information
- Because transcribers make continuous judgment calls about disfluencies, unclear words and accents that must stay consistent
- Because transcription is done entirely by automated software with no humans
- According to the outlook, which conditions are likely to remain hardest for automated pre-labeling of speech?
- Clearly spoken, single-speaker recordings in a well-resourced language
- Any audio that is shorter than one minute
- Recordings made in a quiet studio
- Accents, code-switching between languages, and low-resource languages