引導式測驗 · 中等 等級
Griffin-Lim 譜圖反演測驗
追蹤 Griffin-Lim 的交替幅度和相位更新,識別變換設置,並將近似反演與學習的聲碼器進行比較。
問題 1 的 8
Which missing component does Griffin-Lim estimate when reconstructing a waveform from spectral magnitudes?
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。
- Which missing component does Griffin-Lim estimate when reconstructing a waveform from spectral magnitudes?
- The original words represented by the signal
- The recording's speaker identity
- A phase pattern compatible with a waveform
- The class labels used by a downstream audio model
- After inverse STFT, what does a Griffin-Lim iteration do with the new STFT?
- Sets every phase value to zero before inverse transformation
- Replaces its magnitudes with the target while retaining estimated phase
- Updates target magnitudes from the current waveform while keeping the initial phase fixed
- Stops after inverse STFT without checking a new forward STFT
- Why is reconstruction from magnitude generally approximate?
- Magnitude contains the full complex STFT
- Every spectrogram maps to one exact waveform
- The algorithm uses no Fourier transform
- The original phase information has been discarded
- Which settings should match between analysis and reconstruction?
- FFT size, window, hop and centering conventions
- Only the total number of output samples
- Only the sample rate, regardless of window or hop
- Any transform settings that produce the same array dimensions
- Does increasing iteration count guarantee perceptually better audio?
- 是的,總是
- Yes, if random phase is used
- No; effects depend on representation and settings
- No; iterations never change output
- Which is a common role of Griffin-Lim in a speech project?
- It reconstructs the exact original phase from magnitude alone
- A simple baseline for turning magnitude estimates into audio
- It uses a learned neural vocoder to predict the waveform
- It guarantees a natural-sounding waveform from any magnitude input
- How do learned neural vocoders differ from Griffin-Lim?
- They use no learned parameters and only alternate fixed projections
- They must reuse Griffin-Lim's estimated phase at every step
- They only predict transcript words rather than waveform samples
- They use a trained model to map acoustic representations to waveforms
- What should be done when judging Griffin-Lim output quality?
- Listen and evaluate the task-relevant audio, not only a spectral score
- Assume a smaller spectral error guarantees natural speech
- Use iteration count alone as the quality measure
- Judge every recording from one clean reconstruction example