Aquila Resolve - Grapheme-to-Phoneme Converter

Augmented Recurrent Neural G2P with Inflectional Orthography

Aquila Resolve presents a new approach for accurate and efficient English to ARPAbet G2P resolution. The pipeline employs a context layer, multiple transformer and n-gram morpho-orthographical search layers, and an autoregressive recurrent neural transformer base. The current implementation offers state-of-the-art accuracy for out-of-vocabulary (OOV) words, as well as contextual analysis for correct inferencing of English Heteronyms.

The package is offered in a pre-trained state that is ready for usage as a dependency or in notebook environments. There are no additional resources needed, other than the model checkpoint which is automatically downloaded on the first usage. See Installation for more information.

1. Dynamic Word Mappings based on context:

g2p.convert('I read the book, did you read it?')
# >> '{AY1} {R EH1 D} {DH AH0} {B UH1 K}, {D IH1 D} {Y UW1} {R IY1 D} {IH1 T}?'

g2p.convert('The researcher was to subject the subject to a test.')
# >> '{DH AH0} {R IY1 S ER0 CH ER0} {W AA1 Z} {T UW1} {S AH0 B JH EH1 K T} {DH AH0} {S AH1 B JH IH0 K T} {T UW1} {AH0} {T EH1 S T}.'

	'The subject was told to read. Eight records were read in total.'
Ground Truth	The S AH1 B JH IH0 K T was told to R IY1 D. Eight R EH1 K ER0 D Z were R EH1 D in total.
Aquila Resolve	The S AH1 B JH IH0 K T was told to R IY1 D. Eight R EH1 K ER0 D Z were R EH1 D in total.
Deep Phonemizer (en_us_cmudict_forward.pt)	The S AH B JH EH K T was told to R EH D. Eight R AH K AO R D Z were R EH D in total.
CMUSphinx Seq2Seq (checkpoint)	The S AH1 B JH IH0 K T was told to R IY1 D. Eight R IH0 K AO1 R D Z were R IY1 D in total.
ESpeakNG (with phonecodes)	The S AH1 B JH EH K T was told to R IY1 D. Eight R EH1 K ER0 D Z were R IY1 D in total.

2. Leading Accuracy for unseen words:

g2p.convert('Did you kalpe the Hevinet?')
# >> '{AY1} {R EH1 D} {DH AH0} {B UH1 K}, {D IH1 D} {Y UW1} {R IY1 D} {IH1 T}?'

	"tensorflow"	"agglomerative"	"necrophages"
Aquila Resolve	T EH1 N S ER0 F L OW2	AH0 G L AA1 M ER0 EY2 T IH0 V	N EH1 K R OW0 F EY2 JH IH0 Z
Deep Phonemizer (en_us_cmudict_forward.pt)	T EH N S ER F L OW	AH G L AA M ER AH T IH V	N EH K R OW F EY JH IH Z
CMUSphinx Seq2Seq (checkpoint)	T EH1 N S ER0 L OW0 F	AH0 G L AA1 M ER0 T IH0 V	N AE1 K R AH0 F IH0 JH IH0 Z
ESpeakNG (with phonecodes)	T EH1 N S OW0 R F L OW2	AA G L AA1 M ER0 R AH0 T IH2 V	N EH1 K R AH0 F IH JH EH0 Z

Installation

pip install aquila-resolve

A pre-trained model checkpoint (~106 MB) will be automatically downloaded on the first use of relevant public methods that require inferencing. For example, when instantiating G2p. You can also start this download manually by calling Aquila_Resolve.download().

If you are in an environment where remote file downloads are not possible, you can also transfer the checkpoint manually, placing model.pt within the Aquila_Resolve.data module folder.

Usage

1. Module

from Aquila_Resolve import G2p

g2p = G2p()

g2p.convert('The book costs $5, will you read it?')
# >> '{DH AH0} {B UH1 K} {K AA1 S T S} {F AY1 V} {D AA1 L ER0 Z}, {W IH1 L} {Y UW1} {R IY1 D} {IH1 T}?'

Optional parameters when defining a G2p instance:

Parameter	Default	Description
device	'cpu'	Device for Pytorch inference model. GPU is supported using 'cuda'

Optional parameters when calling convert:

Parameter	Default	Description
process_numbers	True	Toggles conversion of some numbers and symbols to their spoken pronunciation forms. See numbers.py for details on what is covered.

2. Command Line

A simple wrapper for text conversion is available through the aquila-resolve command

~
❯ aquila-resolve
✔ Aquila Resolve v0.1.4
? Text to convert: I read the book, did you read it?
{AY1} {R EH1 D} {DH AH0} {B UH1 K}, {D IH1 D} {Y UW1} {R IY1 D} {IH1 T}?

Model Architecture

In evaluation¹, neural G2P models have traditionally been extremely sensitive to orthographical variations in graphemes. Attention-based mapping of contextual recognition has traditionally been poor for languages like English with a low correlative relationship between grapheme and phonemes². Furthermore, both static methods (i.e. CMU Dictionary), and dynamic methods (i.e. G2p-seq2seq, Phonetisaurus, DeepPhonemizer) incur a loss of sentence context during tokenization for training and inference, and therefore make it impossible to accurately resolve words with multiple pronunciations based on grammatical context (Heteronyms).

This model attempts to address these issues to optimize inference accuracy and run-time speed. The current architecture employs additional natural language analysis steps, including Part-of-speech (POS) tagging, n-gram segmentation, lemmatization searches, and word stem analysis. Some layers are universal for all text, such as POS tagging, while others are activated when deemed required for the requested word. Layer information is retained with the token in vectorized and tensor operations. This allows morphological variations of seen words, such as plurals, possessives, compounds, inflectional stem affixes, and lemma variations to be resolved with near ground-truth level of accuracy. This also improves out-of-vocabulary (OOV) inferencing accuracy, by truncating individual tensor size and characteristics to be closer to seen data.

The inferencing layer is built as an autoregressive implementation of the forward DeepPhonemizer model, as a 4-layer transformer with 256 hidden units. The pre-trained checkpoint for Aquila Resolve is trained using the CMU Dict v0.7b corpus, with 126,456 unique words. The validation dataset was split as a uniform 5% sample of unique words, sorted by grapheme length. The learning rate was linearly increased during the warmup steps, and step-decreased during fine-tuning.

Symbol Set

The 2 letter ARPAbet symbol set is used, with numbered vowel stress markers.

Vowels

Phoneme	Example		Phoneme	Example		Phoneme	Example		Phoneme	Example
AA0	Balm		AW0	Ourself		EY0	Mayday		OY0
AA1	Bot		AW1	Shout		EY1	Mayday		OY1
AA2	Cot		AW2	Outdo		EY2	airfreight		OY2
AE0	Bat		AY0	Ally		IH0	Cooking		UH0
AE1	Fast		AY1	Bias		IH1	Exist		UH1
AE2	Midland		AY2	Alibi		IH2	Outfit		UH2
AH0	Central		EH0	Enroll		IY0	Lady		UW0
AH1	Chunk		EH1	Bless		IY1	Beak		UW1
AH2	Outcome		EH2	Telex		IY2	Turnkey		UW2
AO0	Story		ER0	Chapter		OW0	Reo
AO1	Adore		ER1	Verb		OW1	So
AO2	Blog		ER2	Catcher		OW2	Cargo

License

The code in this project is released under Apache License 2.0.

References

Footnotes

1. r-G2P: Evaluating and Enhancing Robustness of Grapheme to Phoneme Conversion by Controlled noise introducing and Contextual information incorporation ↩

2. OTEANN: Estimating the Transparency of Orthographies with an Artificial Neural Network ↩