Difference between revisions of "Acoustic Analysis"
Keithjohnson (talk | contribs) |
Keithjohnson (talk | contribs) |
||
| Line 10: | Line 10: | ||
The script does the following things: |
The script does the following things: |
||
| − | 1) |
+ | 1) Resample sound file to 16,000 Hz. |
| ⚫ | |||
| ⚫ | |||
| ⚫ | |||
| ⚫ | |||
| ⚫ | |||
| ⚫ | |||
| − | 4) Gives each of these peaks a time-stamped score based on amount of change in waveform relative to neighboring samples |
||
| − | Compute a series of Mel Frequency Spectra: |
+ | '''Compute a series of Mel Frequency Spectra:''' |
| − | + | 4) Now, take Mel frequency spectra from 300Hz to 8000Hz, in non-overlapping 5 ms windows, spanning the interval from start to end times. This is done with the ESPS routine melspec() |
|
| − | + | 5) Compare successive windows, and select the top three candidates with most change in the spectrum. This is done with the ESPS routine diff() |
|
| ⚫ | |||
| + | '''Combine these two acoustic landmarks into a burst score''' |
||
| ⚫ | |||
| − | Calculates a burst score (burst strength) for remaining candidates. Using a linear model trained on the burst locations in TIMIT: |
||
| + | |||
| + | 7) Calculate a burst score (burst strength) for the remaining candidates. This is done using a linear discriminant function trained on bursts in TIMIT. 'b' - the burst score is calcuated: b = -1.814 + 0.618*log(waveform_difference) + 0.003*spectral_difference |
||
| + | |||
8) Selects candidate with highest burst score. |
8) Selects candidate with highest burst score. |
||
Revision as of 11:24, 26 April 2018
Burst detection (burst.py)
This script defines a function burst()
-- input 1) the name of a soundfile, 2) start time (in seconds), and (3) end time of an analysis window where we will look for a burst.
-- output 1) a burst_time (in seconds), and 2) a burst_score which is a measure of how much like a burst the burst is.
The script does the following things:
1) Resample sound file to 16,000 Hz.
With the Audio Waveform:
2) Within the window specified by the start and end times, look at each sample and determines whether it is a peak or valley.
3) Find three biggest valleys in the waveform (corresponding to pressure peaks) within specified time window. The biggest valleys are those that have the largest difference at the valley relative to the adjacent samples. This is the waveform_difference.
Compute a series of Mel Frequency Spectra:
4) Now, take Mel frequency spectra from 300Hz to 8000Hz, in non-overlapping 5 ms windows, spanning the interval from start to end times. This is done with the ESPS routine melspec()
5) Compare successive windows, and select the top three candidates with most change in the spectrum. This is done with the ESPS routine diff()
Combine these two acoustic landmarks into a burst score
6) Compare waveform candidates to spectrum candidates, and keep those where the times align.
7) Calculate a burst score (burst strength) for the remaining candidates. This is done using a linear discriminant function trained on bursts in TIMIT. 'b' - the burst score is calcuated: b = -1.814 + 0.618*log(waveform_difference) + 0.003*spectral_difference
8) Selects candidate with highest burst score.
9) If no burst is detected, returns a burst score of 0 and a burst_time of -1.
