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.
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. |
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 |
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 callingAquila_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 theAquila_Resolve.data
module folder.
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. |
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}?
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.
The 2 letter ARPAbet symbol set is used, with numbered vowel stress markers.
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 |
The code in this project is released under Apache License 2.0.