Techniques for improved audio processing using combinations of clipping engines and acoustic models

1 . A method for efficiently training an acoustic model, comprising:

training an acoustic model over a plurality of training iterations by, at each of the plurality of training iterations:

applying the acoustic model to features extracted from training audio data in order to output a set of acoustic model predictions;

decoding, via a decoder, the training audio data based on the set of acoustic model predictions for the iteration in order to generate decoded text for the iteration, wherein the training audio data is clipped at each iteration based further on the decoded text for the iteration;

applying a language model to at least a set of key term sample data in order to output a set of language model predictions with respect to a plurality of key terms, wherein the key term sample data demonstrates use of the plurality of key terms;

clipping, via a clipping engine, the training audio data into a plurality of clips by inputting the decoded text based on the acoustic model output to the clipping engine and inputting the language model output to the clipping engine; and

tuning the acoustic model via a machine learning algorithm using the plurality of clips.

2 . The method of claim 1 , wherein the set of language model predictions output at each iteration is a first set of language model predictions, further comprising at each of the plurality of training iterations:

generating a second set of language model predictions by applying the language model to the decoded text for the iteration; and

providing the second set of language model predictions as feedback to the decoder.

3 . The method of claim 1 , further comprising:

applying the trained acoustic model to features of an audio data set in order to generate a plurality of acoustic predictions for the audio data set.

4 . The method of claim 3 , further comprising:

applying at least one speech recognition model to the plurality of acoustic predictions for the audio data set.

5 . The method of claim 1 , further comprising:

generating a list of key terms based on the key term sample data, wherein each key term among the list of key terms is a word or combination of words which appears in a plurality of audio samples above a threshold, wherein the language model is configured using the list of key terms.

6 . The method of claim 1 , wherein the acoustic model is tuned based further on target domain transcript data, wherein the target domain transcript data is in a target domain to be used by the acoustic model.

7 . The method of claim 1 , further comprising, at each of the plurality of training iterations:

extracting a plurality of features from the training audio data by performing signal processing on the training audio data in order to transform audio waveforms into vectors.

8 . The method of claim 7 , further comprising:

normalizing the audio waveforms, wherein the normalized audio waveforms are transformed into the vectors.

9 . The method of claim 5 , wherein the list of key terms is unique to a single entity.

10 . A non-transitory computer readable medium having stored thereon instructions for causing a processing circuitry to execute a process, the process comprising:

training an acoustic model over a plurality of training iterations by, at each of the plurality of training iterations:

applying the acoustic model to features extracted from training audio data in order to output a set of acoustic model predictions;

decoding, via a decoder, the training audio data based on the set of acoustic model predictions for the iteration in order to generate decoded text for the iteration, wherein the training audio data is clipped at each iteration based further on the decoded text for the iteration;

applying a language model to at least a set of key term sample data in order to output a set of language model predictions with respect to a plurality of key terms, wherein the key term sample data demonstrates use of the plurality of key terms;

clipping, via a clipping engine, the training audio data into a plurality of clips by inputting the decoded text based on the acoustic model output to the clipping engine and inputting the language model output to the clipping engine; and

tuning the acoustic model via a machine learning algorithm using the plurality of clips.

11 . A system for efficiently training an acoustic model, comprising:

a processing circuitry; and

a memory, the memory containing instructions that, when executed by the processing circuitry, configure the system to:

train an acoustic model over a plurality of training iterations wherein, at each of the plurality of training iterations, the system is configured to:

apply the acoustic model to features extracted from training audio data in order to output a set of acoustic model predictions;

decode, via a decoder, the training audio data based on the set of acoustic model predictions for the iteration in order to generate decoded text for the iteration, wherein the training audio data is clipped at each iteration based further on the decoded text for the iteration;

apply a language model to at least a set of key term sample data in order to output a set of language model predictions with respect to a plurality of key terms, wherein the key term sample data demonstrates use of the plurality of key terms;

clip, via a clipping engine, the training audio data into a plurality of clips by inputting the decoded text based on the acoustic model output to the clipping engine and the language model output to the clipping engine; and

tune the acoustic model via a machine learning algorithm using the plurality of clips.

12 . The system of claim 11 , wherein the set of language model predictions output at each iteration is a first set of language model predictions, wherein the system is further configured to, at each of the plurality of training iterations:

generate a second set of language model predictions by applying the language model to the decoded text for the iteration; and

provide the second set of language model predictions as feedback to the decoder.

13 . The system of claim 11 , wherein the system is further configured to:

apply the trained acoustic model to features of an audio data set in order to generate a plurality of acoustic predictions for the audio data set.

14 . The system of claim 13 , wherein the system is further configured to:

apply at least one speech recognition model to the plurality of acoustic predictions for the audio data set.

15 . The system of claim 11 , wherein the system is further configured to:

generate a list of key terms based on the key term sample data, wherein each key term among the list of key terms is a word or combination of words which appears in a plurality of audio samples above a threshold, wherein the language model is configured using the list of key terms.

16 . The system of claim 11 , wherein the acoustic model is tuned based further on target domain transcript data, wherein the target domain transcript data is in a target domain to be used by the acoustic model.

17 . The system of claim 11 , wherein the system is further configured to, at each of the plurality of training iterations:

extract a plurality of features from the training audio data by performing signal processing on the training audio data in order to transform audio waveforms into vectors.

18 . The system of claim 17 , wherein the system is further configured to:

normalize the audio waveforms, wherein the normalized audio waveforms are transformed into the vectors.