System and method for compressed domain language identification
View Patent ↗Embodiments included herein are directed towards a system and method for compressed domain language identification. Embodiments may include receiving a bitstream of a sequence of packets at one or more computing devices and classifying each packet into speech or non-speech based upon, at least in part, compressed domain voice activity detection (VAD). Embodiments may further include extracting a pseudo-cepstral representation from the speech detected packets and partially decoding without extracting a PCM format and generating a sequence of multi-frames, based upon, at least in part, the pseudo-cepstral representation. Embodiments may also include providing in real time the sequence of multi-frames to a deep neural network (DNN), wherein the DNN has been trained off-line for one or more desired target languages.
1. A compressed domain language identification method comprising:
receiving a bitstream of a sequence of packets at one or more computing devices;
classifying each packet into speech or non-speech based upon, at least in part, compressed domain voice activity detection (VAD);
extracting a pseudo-cepstral representation from the speech detected packets and partially decoding without extracting a PCM format;
generating a sequence of multi-frames, based upon, at least in part, the pseudo-cepstral representation; and
providing in real time the sequence of multi-frames to a deep neural network (DNN), wherein the DNN has been trained off-line for one or more desired target languages.
2. The method of claim 1 , further comprising:
automatically detecting a language associated with the bitstream of the sequence of packets.
3. The method of claim 1 , wherein extracting includes extracting from a non-decoded bitstream of the sequence of packets.
4. The method of claim 1 , wherein extracting includes a delta computation and a delta-delta computation.
5. The method of claim 1 , further comprising:
discarding one or more non-speech frames using a compressed domain voice activity detector (VAD).
6. The method of claim 1 , wherein the DNN is configured to classify the bitstream as belonging to a particular language.
7. The method of claim 6 , further comprising:
averaging a logarithm of posterior possibilities of the DNN across one or more packets to detect the particular language.
8. The method of claim 1 , further comprising:
detecting a language associated with the bitstream of the sequence of packets without decoding into a pulse code modulated format.
9. The method of claim 1 , further comprising:
receiving an output score from the DNN; and
combining the output score with a pulse code modulated (PCM) language identification system score.
10. The method of claim 6 , further comprising:
performing an action based upon the classification, wherein the action includes at least one of, automatically routing a call in a call center, non-intrusively obtaining statistics regarding language usage in a network, and performing speaker identification.
11. A system for compressed domain language identification, the system including at least one processor configured to perform operations comprising:
receiving a bitstream of a sequence of packets at one or more computing devices;
classifying each packet into speech or non-speech based upon, at least in part, compressed domain voice activity detection (VAD);
extracting a pseudo-cepstral representation from the speech detected packets and partially decoding without extracting a PCM format;
generating a sequence of multi-frames, based upon, at least in part, the pseudo-cepstral representation; and
providing in real time the sequence of multi-frames to a deep neural network (DNN), wherein the DNN has been trained off-line for one or more desired target languages.
12. The system of claim 11 , further comprising:
automatically detecting a language associated with the bitstream of the sequence of packets.
13. The system of claim 11 , wherein extracting includes extracting from a non-decoded bitstream of the sequence of packets.
14. The system of claim 11 , wherein extracting includes a delta computation and a delta-delta computation.
15. The system of claim 11 , further comprising:
discarding one or more non-speech frames using a compressed domain voice activity detector (VAD).
16. The system of claim 11 , wherein the DNN is configured to classify the bitstream as belonging to a particular language.
17. The system of claim 16 , further comprising:
averaging a logarithm of posterior possibilities of the DNN across one or more packets to detect the particular language.
18. The system of claim 11 , further comprising:
detecting a language associated with the bitstream of the sequence of packets without decoding into a pulse code modulated format.
19. The system of claim 11 , further comprising:
receiving an output score from the DNN; and
combining the output score with a pulse code modulated (PCM) language identification system score.
20. The system of claim 16 , further comprising:
performing an action based upon the classification, wherein the action includes at least one of, automatically routing a call in a call center, non-intrusively obtaining statistics regarding language usage in a network, and performing speaker identification.