Hot answers tagged

14

This is a well-studied problem, dating back from the mid 90s (DARPA/NIST broadcast transcription challenges). Search for "speech/music segmentation" or "audio segmentation" and you'll find thousands of research papers. There are two broad approaches to solve this problem: Supervised classification Train a speech/music classifier, using a standard machine ...


11

What you want to do is called "Super Resolution" in the field of imaging. This is an ill posed problem. Hence in order to solve it you need some prior / model about your audio data. For example you may look at the model in Speech Super Resolution Using Parallel WaveNet.


8

Now, I would like to show what frequencies the speech has. However, I'm not sure what would be the best way to do that. It seems sometimes one calculates the absolute value of a Fourier transform, and sometimes power spectral density. If you want to attach physical meaning to your analysis, then go with the power spectral density, (PSD). This is ...


8

While I don't know about tools for upsampling using AI, I am challenging the assumption that the main problem with your sample is lost high frequncies due to resampling from 44kHz to 22kHz, so "just" upsampling is unlikely to solve your root problem. The possibility of AI upsampling itself is sensible, though: While standard (mathematical) ...


6

A recording originally at 8kHz and digitally upsampled to 16kHz will have almost no energy in the 4-8kHz range (whatever is here is due to imperfections in the filters used for the upsampling process). I would just use a 4kHz and 5.5kHz high pass; and use a threshold on the signal energy at the output of these filters. ... Unless your recordings are ...


6

Researchers from the Johns Hopkins University have recently released a corpus of music, speech, and noise which, according to them, is suitable for training models for voice activity detection and music/speech discrimination. See https://arxiv.org/pdf/1510.08484.pdf for details.


6

The i-vectors and x-vectors share the ability to represent speech utterance in a compact way (as a vector of fixed size, regardless of length of the utterance). The extraction algorithms of i-vectors and x-vector are quite different. The x-vector concept is newer and the name of the method is similar to "i-vector" to suggests that this representation can be ...


5

The amount of knowledge necessary to develop such a large scale multi-language, speaker-independent, large-vocabulary speech recognition system is spread well over hundreds of papers; and each individual brick of the system (say the feature extraction front-end, the FST decoding library, the language model store) is developed by world-class expert in this ...


5

Depending on the actual recordings, the algorithm complexity could range from dead easy to really complex... I'll take the studio recording case first, so I can assume : - (Almost) no noise coming from outside (cars, trucks, bus...) - Nobody slamming the door in the middle of the recording - Voice samples are recorded at optimal level independently of who ...


4

Assume noise is not a serious issue in your problem. I guess you can get pretty clear speech signal. If you have speech recognition part implemented in your system, I think you should be able to take advantage over the language model in your recognition system. According to the transition probabilities, you shall get some confidence to say at what moment ...


4

White noise implies no correlation between samples of the noise, even consecutive samples. Colored noise, therefore, implies that there is correlation of some sort between the noise samples, which in turn implies that we can take advantage of that correlation to get rid of some of the noise. Beyond that, there is not a lot that we can say about what it ...


4

Voice Activity Detection using Adaptive Threshold is very easy and handy to implement on any platform Here you can have a algorithm which is Adaptive Energy based Small addition to above algorithm when you are calculating for very first time go for taking Mean of Energy and mark as Emin in this the frame you pass is divided in to sub-frames and further ...


4

It's better to copy first frame and last frame values to extend vector sequence beyond boundaries than to assign 0. This could be implemented just by adjusting indexes: if (index1 < 0) index1 = 0 if (index2 > N - 1) index2 = N - 1 delta = v[index1] - v[index2]


4

You can take some big speech corpus like TEDLIUM and add the noise you like: http://www-lium.univ-lemans.fr/en/content/ted-lium-corpus The advantage of TEDLIUM is that it's a set of continuous recordings with speech timings, not just a collection of utterances.


4

Dynamic Time Warping is pretty well explained on this site. I'll use some of the diagrams from the PPT on that site to explain. The idea is to divide the signals into segments (frames) and then compare frames sequentially through each signal. As illustrated below, motion from a segment in one signal to the next segment depends on the similarity to the ...


4

What you are observing is the digital representation of the voltage, which in fact represents the acoustic pressure. Workflow would be something like: Vibrating larynx is producing Acoustic Pressure [Pa] Variations of that pressure are converted by the microphone into Voltage [V] Voltage is being sampled and quantized by the ADC (Analog to Digital Converter)...


4

Yes, the cellular phones use various forms of compression to convert the captured analog audio (speech) into the digital bitstream for transmission through 2G/3G/.../. The specific method used depends on the GSM version which might dictate its own bandwidth constraint and backward or forward compatibility issues into the audio encoding stage. Most ...


4

Band-pass filtering with cut-off frequencies of 300 Hz and 3400 Hz should result in a good approximation. Try with a Chebychev filter or order not more than 6. Then you may need to downsample your audio to 8000 samples per second, which is the standard for telephony. P.S. The actual cut-off frequencies (especially the 3400 Hz) may be different according to ...


4

On each individual device, the speaker output can get subtracted from the microphone before it gets sent to other locations. This prevents others from hearing themselves through your microphone. When using two devices in within audible range of each other, the devices cannot subtract the speaker audio from the microphone audio because the information path ...


4

As pointed in some of the comments, there are several issues involved in your problem that you should consider to go for an optimal solution: Quality of the record system. Features of the microphone, filter of the preamps, etc. What is the frequency response of the system (mic + preamp + adc)? https://blog.faberacoustical.com/2009/ios/iphone/iphone-...


3

Yes, this should be enough for a basic isolated word recognition system. Probably not something for a commercial product, but good enough for a university project or demo... It would be better to ask the user to record a word and match against this, rather than attempt to match against a large database of utterances of the same word by different speakers. ...


3

The important thing to understand about something like a speech signal is that its frequency components are time-varying. In order to represent speech in the frequency domain we usually take a short enough window of the signal within which we can assume that the spectrum of the speech does not vary significantly (typically 10 ms). So we calculate the power ...


3

There is no best pdf for a periodic signal. There is also no way to find the 'exact' pdf of a measured signal. What you have to do is to measure the pdf from the data. Use a histogram to approximate the pdf of your data. Define a number of intervals within the amplitude range of your signal and simply count the numbers of data samples per interval. This ...


3

I think you have a misunderstanding about formants, pitch and partial tones. The spectrogram shows the position of the partials and you can also estimate their fundamental frequency, or pitch, trajectory to be the position of the lowest partial. The formant structure is not readily recognizable from this representation. Formants are the resonances of the ...


3

How about LibVAD? Seems like that does exactly what you're describing. Disclosure: I'm the developer behind LibVAD


3

Audio watermarking is a technique that relies on inaudible cues. Humans can't really hear phase differences in low tones, if only because it's often delivered via a single bass speaker (i.e. because the HiFi setup physically eliminates the difference). This allows you to stuff information in the phases of the low-frequency parts of the left and right channel....


3

I just came across one article related to the compand command that might be useful (just in case someone looking for help sees this article). From the article: To test this, and my understanding of compand, I added a very simple filter to remove the quiet section of the audio by decreasing their volume dramatically: ffmpeg -i in.mp3 -filter_complex \ ...


3

The frequency resolution of DFT is limited to the number of time samples. On the other hand proper LPC can have high resolution.


3

I will not address the question whether or not it is meaningful to compute the instantaneous frequency of a speech signal. Instead I will show you a better method for computing the instantaneous frequency from a given analytic signal. This method avoids the phase unwrapping problem by directly computing the instantaneous frequency from the real and imaginary ...


3

Ok so here is what I found. The distance is dependent on the way that the mfccs are calculated. This makes sense to me and also explains why cepstral mean normalization affects the values of the MCD. I found this implementation (https://github.com/MattShannon/mcd), which unfortunately did not support .wav files. I ran this and it gives results that are in ...


Only top voted, non community-wiki answers of a minimum length are eligible