Large-scale language model data selection for rare-word speech recognition
A method of training a language model for rare-word speech recognition includes obtaining a set of training text samples, and obtaining a set of training utterances used for training a speech recognition model. Each training utterance in the plurality of training utterances includes audio data corresponding to an utterance and a corresponding transcription of the utterance. The method also includes applying rare word filtering on the set of training text samples to identify a subset of rare-word training text samples that include words that do not appear in the transcriptions from the set of training utterances or appear in the transcriptions from the set of training utterances less than a threshold number of times. The method further includes training the external language model on the transcriptions from the set of training utterances and the identified subset of rare-word training text samples.
1. A computer-implemented method executing on data processing hardware that causes the data processing hardware to perform operations comprising:
receiving a corpus of training text samples, each training text sample in the corpus of training text samples comprising a corresponding sentence;
executing a resampling function on the corpus of training text samples that downsamples a frequency distribution of the corpus of training text samples by:
matching the frequency distribution of the corpus of training text samples up to a threshold frequency; and
applying logarithmic scaling on the frequency distribution of the corpus of training text samples after the threshold frequency to identify high frequency training text samples as the training text samples from the corpus of training text samples that have corresponding frequencies exceeding the threshold frequency;
obtaining a set of training text samples by removing the identified high frequency training text samples from the corpus of training text samples;
applying rare word filtering on the training text samples to identify a subset of rare-word training text samples; and
training a language model on the identified subset of rare-word training text samples.
2. The computer-implemented method of claim 1 , wherein the operations further comprise determining the frequency distribution of the corpus of training text samples that identifies a corresponding frequency that each training text sample in the corpus of training text samples occurs relative to the corresponding frequencies of the other training text samples in the corpus of training text samples.
3. The computer-implemented method of claim 1 , wherein the resampling function comprises one of a simple power resampling function, a forced power resampling function, or a soft logarithmic resampling function.
4. The computer-implemented method of claim 1 , wherein the operations further comprise:
applying contrastive filtering on the set of training text samples to identify a subset of target domain training text samples that match a target domain associated with the set of training text samples,
wherein training the language model on the identified subset of rare-word training text samples further comprises training the language model on the identified subset of target domain training text samples that match the target domain.
5. The computer-implemented method of claim 1 , wherein the language model comprises a neural language model.
6. The computer-implemented method of claim 5 , wherein the neural language model comprises a stack of conformer layers or transformer layers.
7. The computer-implemented method of claim 1 , wherein the operations further comprise:
obtaining a set of training utterances used for training a machine learning model, each training utterance comprising a corresponding transcription of an utterance,
wherein the identified subset of are rare training text samples includes words that do not appear in the transcriptions from the set of training utterances or appear in the transcriptions from the set of training utterances less than a threshold number of times.
8. The computer-implemented method of claim 7 , wherein:
the machine learning model comprises a machine translation model; and
training the language model on the identified subset of rare-word training text samples further comprises training the language model on the transcriptions from the set of training utterances.
9. The computer-implemented method of claim 7 , wherein:
the machine learning model comprises an automatic speech recognition (ASR) model;
each training utterance in the set of training utterances used for training the ASR model further comprises audio data corresponding to the utterance; and
training the language model on the identified subset of rare-word training text samples further comprises training the language model on the transcriptions from the set of training utterances.
10. The computer-implemented method of claim 9 , wherein the operations further comprise integrating a trained external language model with the trained ASR model, the trained external language model configured to rescore probability distributions over possible speech recognition hypotheses predicted by the trained ASR model.
11. A system comprising
data processing hardware; and
memory hardware storing instructions that when executed on the data processing hardware causes the data processing hardware to perform operations comprising:
receiving a corpus of training text samples, each training text sample in the corpus of training text samples comprising a corresponding sentence;
executing a resampling function on the corpus of training text samples that downsamples a frequency distribution of the corpus of training text samples by:
matching the frequency distribution of the corpus of training text samples up to a threshold frequency; and
applying logarithmic scaling on the frequency distribution of the corpus of training text samples after the threshold frequency to identify high frequency training text samples as the training text samples from the corpus of training text samples that have corresponding frequencies exceeding the threshold frequency;
obtaining a set of training text samples by removing the identified high frequency training text samples from the corpus of training text samples;
applying rare word filtering on the training text samples to identify a subset of rare-word training text samples; and
training a language model on the identified subset of rare-word training text samples.
12. The system of claim 11 , wherein the operations further comprise determining the frequency distribution of the corpus of training text samples that identifies a corresponding frequency that each training text sample in the corpus of training text samples occurs relative to the corresponding frequencies of the other training text samples in the corpus of training text samples.
13. The system of claim 11 , wherein the resampling function comprises one of a simple power resampling function, a forced power resampling function, or a soft logarithmic resampling function.
14. The system of claim 11 , wherein the operations further comprise:
applying contrastive filtering on the set of training text samples to identify a subset of target domain training text samples that match a target domain associated with the set of training text samples,
wherein training the language model on the identified subset of rare-word training text samples further comprises training the language model on the identified subset of target domain training text samples that match the target domain.
15. The system of claim 11 , wherein the language model comprises a neural language model.
16. The system of claim 15 , wherein the neural language model comprises a stack of conformer layers or transformer layers.
17. The system of claim 11 , wherein the operations further comprise:
obtaining a set of training utterances used for training a machine learning model, each training utterance comprising a corresponding transcription of an utterance,
wherein the identified subset of rare-word training text samples includes words that do not appear in the transcriptions from the set of training utterances or appear in the transcriptions from the set of training utterances less than a threshold number of times.
18. The system of claim 17 , wherein:
the machine learning model comprises a machine translation model; and
training the language model on the identified subset of rare-word training text samples further comprises training the language model on the transcriptions from the set of training utterances.
19. The system of claim 17 , wherein:
the machine learning model comprises an automatic speech recognition (ASR) model;
each training utterance in the set of training utterances used for training the ASR model further comprises audio data corresponding to the utterance; and
training the language model on the identified subset of rare-word training text samples further comprises training the language model on the transcriptions from the set of training utterances.
20. The system of claim 19 , wherein the operations further comprise integrating a trained external language model with the trained ASR model, the trained external language model configured to rescore probability distributions over possible speech recognition hypotheses predicted by the trained ASR model.