Method and system of audio input bit-size conversion for audio processing

Patent metadata

Field	Value
Publication number	US-11875783-B2
Application number	US-202016892080-A
Country	US
Kind code	B2
Filing date	Jun 3, 2020
Priority date	Jun 3, 2020
Publication date	Jan 16, 2024
Grant date	Jan 16, 2024

Abstract

A method, system, and device are directed to audio input bit-size conversion for compatibility to audio processing systems with an expected input sample bit-size.

First claim

What is claimed is:

1. An audio processing device comprising:
memory storing audio input including human speech and in a form of initial samples with a first bit-size; and
at least one processor communicatively coupled to the memory to operate by:
dividing at least one of the initial samples into multiple sample parts;
generating at least one gain formed by at least one neural network accelerator;
applying the at least one gain to at least one of the sample parts to form at least one scaled sample part; and
generating a scaled output sample in a second bit-size comprising combining at least portions of the multiple sample parts including the at least one scaled sample part, and wherein a portion of one of the sample parts being combined has most significant bits (MSBs) of the initial sample and a portion of another one of the sample parts being combined has least significant bits (LSBs) of the initial sample.

2. The device of claim 1 , wherein the sample parts each have a size so that the sample parts cooperatively hold all of the bits from the initial sample.

3. The device of claim 1 , wherein the sample parts are of the second bit-size.

4. The device of claim 1 , wherein the sample parts comprise at least a high sample part filled with the most significant bits and other bits from the initial sample and a low sample part having the least significant bits from the initial sample and remaining bit spaces filled with zeros.

5. The device of claim 1 , wherein the dividing comprises storing the initial sample in a container of a transition sample with a third bit-size that is larger than the first bit-size of the initial sample and evenly divisible into the sample parts.

6. The device of claim 5 , wherein the first bit-size is 24 bits, the second bit-size is 16 bits, and the third bit-size is 32 bits.

7. The device of claim 5 , wherein the at least one processor is arranged to operate by deinterleaving a sequence of the transition samples, wherein each transition sample has a high sample part and a low sample part, and the deinterleaving to generate a high sample vector of high sample parts separate from a low sample vector of low sample parts to separately input the high and low sample vectors into a neural network accelerator.

8. The device of claim 7 , wherein the at least one processor to shift the low sample parts having the least significant bits (LSBs) of the initial samples to reserve a bit space in the low sample part for a sign bit using at least one neural network accelerator.

9. The device of claim h wherein the at least one processor operates by determining absolute value versions of the sample parts and a separate sign vector maintaining a sign of at least one of the sample parts to use to generate the scaled output sample.

10. A method of audio processing comprising:
obtaining audio input including human speech and in a form of initial samples with a first bit-size;
dividing at least one of the initial samples into multiple sample parts;
generating, by at least one neural network accelerator, at least one gain;
applying the at least one gain to at least one of the sample parts to form at least one scaled sample part; and
generating a scaled output sample in a second bit-size comprising combining at least portions of the multiple sample parts and including the at least one scaled sample part, and wherein a portion of one of the sample parts being combined has most significant bits (MSBs) of the initial sample and a portion of another one of the sample parts being combined has least significant bits (LSBs) of the initial sample.

11. The method of claim 10 , wherein the at least one gain is computed dynamically depending on the sample parts.

12. The method of claim 10 , wherein the at least one gain is computed by using a count of a number of bit spaces occupied by one of the sample parts.

13. The method of claim 10 , wherein the same at least one gain is used for multiple sample parts of a same sample set of multiple parts of multiple initial samples regardless of which sample part was used to form the gain.

14. The method of claim 10 , wherein multiple initial samples of a sample set of initial samples are divided into sample parts, and wherein the at least one gain is generated by using only data of a high sample part with the highest value among all high sample parts of the set.

15. The method of claim 14 , comprising determining the high sample part with the highest value by using max pooling layers of a neural network.

16. A computer-implemented system for audio processing comprising:
at least one microphone to capture audio input including human speech;
memory to store the audio input in of initial samples of a first bit-size;
at least one processor communicatively coupled to the at least one microphone and at least one memory, and to operate by:
dividing at least one of the initial samples into multiple sample parts;
generating at least one gain formed by at least one neural network accelerator;
applying the at least one gain to at least one of the sample parts to form at least one scaled sample part; and
generating a scaled output sample in a second bit-size comprising combining at least portions of the multiple sample parts and including the at least one scaled sample part, and wherein a portion of one of the sample parts being combined has most significant bits (MSBs) of the initial sample and a portion of another one of the sample parts being combined has least significant bits (LSBs) of the initial sample.

17. The system of claim 16 , wherein the at least one gain is arranged so that applying the at least one gain causes a bit-shift in the sample part to place a most significant bit of the sample part at the highest available bit space of a scaled sample part to be used to form the scaled output sample.

18. The system of claim 17 , wherein the bit-shift provides empty bit spaces on the scaled sample part to receive bits of a scaled low sample part associated with the least significant bits of the initial sample.

19. The system of claim 16 , wherein the scaled output sample is formed by combining at least portions of a scaled high sample part and a scaled low sample part.

20. At least one non-transitory machine readable medium comprising instructions that, in response to being executed on a computing device, cause the computing device to operate by:
obtaining audio input including human speech and in a form of initial samples with a first bit-size;
dividing at least one of the initial samples into multiple sample parts;
generating, by at least one neural network accelerator, at least one gain;
applying the at least one gain to at least one of the sample parts to form at least one scaled sample part; and
generating a scaled output sample in a second bit-size comprising combining at least portions of the multiple sample parts and including the at least one scaled sample part, and wherein a portion of one of the sample parts being combined has most significant bits (MSBs) of the initial sample and a portion of another one of the sample parts being combined has least significant bits (LSBs) of the initial sample.

21. The machine readable medium of claim 20 , wherein at least one of the dividing, applying the at least one gain, and generating a scaled output sample are performed by one or more neural network accelerators without the use of a digital signal processor (DSP).

22. The machine readable medium of claim 20 , wherein the instructi

Assignees

• Intel Corp

Inventors

• Terpilowski Marcin
• Dorau Tomasz
• Bocklet Tobias

Classifications

• G10L15/16 Primary

  using artificial neural networks · CPC title

• G06F3/162

  Interface to dedicated audio devices, e.g. audio drivers, interface to CODECs · CPC title

• G06F3/165

  Management of the audio stream, e.g. setting of volume, audio stream path · CPC title

• H03M7/28 Primary

  Programmable structures, i.e. where the code converter contains apparatus which is operator-changeable to modify the conversion process · CPC title

• H03M7/04

  the radix thereof being two · CPC title