This is a list of existing g2p repos, directions on how to get the respective models, and how to use these models to get phonemes.

Home

Table of Contents

• Specifications
• grammatek/g2p-lstm
• cadia-lvl/tts-data
• atliSig/g2p
• rkjaran/g2p-service
• grammatek/g2p-thrax
• kaldi/egs/althingi/s5
• template

Specifications

Repo name	Toolkit	IPA or SAMPA	PER*	Details
g2p-lstm	fairseq	SAMPA	4.07% (standard)	grammatek/g2p-lstm
tts-data	sequitur	IPA	N/A	cadia-lvl/tts-data
g2p	sequitur	IPA	N/A	atliSig/g2p
g2p-service	sequitur	NA/IPA	N/A	rkjaran/g2p-service
g2p-thrax	thrax	SAMPA	6.49%	grammatek/g2p-thrax
althingi	sequitur	IPA	N/A	kaldi/egs/althingi/s5

* This PER is based on the Icelandic Language Technology g2p test set with an emphasis on good coverage of grapheme combinations so this is higher than it would be in a normal texts.

grammatek/g2p-lstm

Type:

fairseq lstm (python) models SAMPA

Location:

the models are available in checkpoints and data-bin

There are g2p models for Icelandic four main regional dialects: north, northest, south, and standard(west).

Dependencies:

Install all dependencies with the following conda env create -f environment.yml

Usage Example

Here’s a usage example on the command line

$ cat example.txt
h l a u p a
d e r p
$ cat example.txt | ./transcribe_ice_standard \
 example_g2p_standard.out example_g2p_standard.tsv
$ cat example_g2p_standard.tsv
h l a u p a	l_0 9i: p a
d e r p	t E r_0 p

cadia-lvl/tts_data

Type:

Sequitur model IPA

Location:

model available here: pron_data/ipd_clean_slt2018.mdl

Dependencies: just do pip3 install -r requirements

Usage:

tests/tests.py gives an example of to how to call it.

Example

input file: <phrase>tab<list-name>

Mammon vor er alhreinn orðinn	althingi
forseti	althingi

output file: <phrase>tab<list-name>tab<phonemes>

Mammon vor er alhreinn orðinn	althingi	~ m a m ɔ n	v ɔː r	ɛ r	aː l̥ r̥ ei t n̥	ɔ r ð ɪ n ~
forseti	althingi	~ f ɔ r̥ s ɛ t ɪ ~

atliSig/g2p

This model was used for creating tts data collection scripts and for training unit selection voices by LVL.

Type

Sequitur model IPA

Location

https://github.com/atliSig/g2p/blob/master/data/ipd_clean_slt2018.mdl

Dependencies:

They’re all listed in requirements.txt within the repo

Usage

    Input arguments:
    * src_path (str): The path to the file containing multiple
    utterances, one per line
    * out_path (str): The target path the the file that stores
    the results.
    * n_jobs (int): The maximum number of processes that can
    be used to execute the given calls.
    * contains_scores (bool): If True, each line in the input file
    is e.g. <sentence>\t<source_id>\t<score> else it is
    <sentence>\t<source_id>
    * translator_options (Options instance): Options passed onto g2p.Translator

Example

input - tsv file

Ég held að það sé ekki rétt orð, þetta komi ekkert almanakinu við.
Fréttirnar eru sko aldeilis dæmalausar.
Það var miði í póstkassanum.
Hann glennti fætur hennar í sundur og reyndi að þrengja sér inn í hárlaust skaut hennar.
Ég tel að þetta sé óskaplega mikil sóun á mannauði hjá einni þjóð.
Yfirverkfræðingur hjá Reykjavíkurborg segir að sökin liggi hjá ökumönnum.
Og nú átti sem sagt að hafa sama háttinn á gagnvart stúdentum hér við Háskólann.
Bæði almennt og sérstakt veiðigjald er lagt flatt sem föst krónutala á þorskígildiskíló.
Lesendur þurfa að fá tækifæri til að njóta hans.

output - tsv file

Ég held að það sé ekki rétt orð, þetta komi ekkert almanakinu við.	~ j ɛː ɣ	h ɛ l t	aː ð	θ aː ð	s j ɛː	ɛ h c ɪ	r j ɛ h t	ɔ r ð	θ ɛ h t a	kʰ ɔː m ɪ	ɛ h c ɛ r̥ t	a l m a n aː c ɪ n ʏ	v ɪː ð ~
Fréttirnar eru sko aldeilis dæmalausar.	~ f r j ɛ h t ɪ t n a r	ɛː r ʏ	s k ɔ	a l t ei l ɪ s	t aiː m a l œyː s a r ~
Það var miði í póstkassanum.	~ θ aː ð	v aː r	m ɪː ð ɪ	iː	pʰ ou s t k a s a n ʏ m ~
Hann glennti fætur hennar í sundur og reyndi að þrengja sér inn í hárlaust skaut hennar.	~ h a n	k l ɛ n̥ t ɪ	f aiː t ʏ r	h ɛ n a r	iː	s ʏ n t ʏ r	ɔː ɣ	r ei n t ɪ	aː ð	θ r ei ɲ c a	s j ɛː r	ɪ n	iː	h au r l œy s t	s k œyː t	h ɛ n a r ~
Ég tel að þetta sé óskaplega mikil sóun á mannauði hjá einni þjóð.	~ j ɛː ɣ	tʰ ɛː l	aː ð	θ ɛ h t a	s j ɛː	ouː s k a p l ɛ ɣ a	m ɪː c ɪ l	s ouː ʏ n	auː	m a n œy ð ɪ	ç au	ei t n ɪ	θ j ouː ð ~
Yfirverkfræðingur hjá Reykjavíkurborg segir að sökin liggi hjá ökumönnum.	~ ɪː v ɪ r v ɛ r̥ k f r ai ð i ŋ k ʏ r	ç au	r eiː c a v iː k ʏ r p ɔ r k	s ei j ɪ r	aː ð	s œː c ɪ n	l ɪ c ɪ	ç au	œː k ʏ m œ n ʏ m ~
Og nú átti sem sagt að hafa sama háttinn á gagnvart stúdentum hér við Háskólann.	~ ɔː ɣ	n uː	au h t ɪ	s ɛː m	s a x t	aː ð	h aː v a	s aː m a	h au h t ɪ n	auː	k a k v a r̥ t	s t uː t ɛ n̥ t ʏ m	ç ɛː r	v ɪː ð	h auː s k ou l a n ~
Bæði almennt og sérstakt veiðigjald er lagt flatt sem föst krónutala á þorskígildiskíló.	~ p aiː ð ɪ	a l m ɛ n̥ t	ɔː ɣ	s j ɛ r̥ s t a x t	v eiː ð ɪ c a l t	ɛ r	l a x t	f l a h t	s ɛː m	f œ s t	kʰ r ouː n ʏ tʰ aː l a	auː	θ ɔ r̥ s c i c ɪ l t ɪ s c i l ou ~
Lesendur þurfa að fá tækifæri til að njóta hans.	~ l ɛː s ɛ n t ʏ r	θ ʏ r v a	aː ð	f auː	tʰ aiː c ɪ f aiː r ɪ	tʰ ɪː l	aː ð	n j ouː t a	h a n s ~

rkjaran/g2p-service

Type

sequitur wrapper

The endpoint is IPA but the repo is just whatever model type you put in.

Location

It doesn’t come with a model but there’s an endpoint: https://nlp.talgreinir.is/pron/derp

Dependencies

Install via dockerfile

Usage

Example

single word

$ curl -XGET https://nlp.talgreinir.is/pron/derp | jq
[
  {
    "results": [
      {
        "posterior": 0.9138450652404999,
        "pronunciation": "t ɛ r̥ p"
      }
    ],
    "word": "derp"
  }
]

Multiple word support with a POST.

$ cat <<EOF | curl -XPOST -d@- https://nlp.talgreinir.is/pron | jq
{"words": ["herp", "derp"]}
EOF
[
  {
    "results": [
      {
        "posterior": 0.9251423160703962,
        "pronunciation": "h ɛ r̥ p"
      }
    ],
    "word": "herp"
  },
  {
    "results": [
      {
        "posterior": 0.9138450652404999,
        "pronunciation": "t ɛ r̥ p"
      }
    ],
    "word": "derp"
  }
]

Append ?t=tsv to get the response in the Kaldi lexicon format.

Hasler 1.0 h a s t l ɛ r

grammatek/g2p-thrax

Type

thrax - rule based SAMPA

Location

need to build it yourself from the repo

Dependencies:

Thrax and OpenFST

Usage

It’s a good idea to put thraxg2p in your path so you can use it from anywhere. This uses x_sampa which means it’s a one to one mapping of IPA to a ASCII only transcription

Example

If you are getting the phonemes interactively from the project root directory,

$./build/thraxg2p --far=grammars/g2p.far
Input string: orð
Output string:  O r D

You can also pipe it in through stdin.

Doing file input and output

./build/thraxg2p --far=grammars/g2p.far --indir=temp \
 --word_file=temp/test.txt > temp/out/test.tsv

input file

derp
orð

output file

derp	t E r_0 p
orð	O r D

althingi/s5

Type

sequitur IPA

Location

There are no public models but you can contact lvl.ru.is if you would like a copy.

terra:/models/althingi

Dependencies:

sequitur-g2p

Usage

It’s best to use this only for ASR, not TTS because it’s not precise enough.

train a g2p model with local/train_g2p.sh

transcribe words with local/transcribe_g2p.sh

Example

transcribing words

$ srun ./local/transcribe_g2p.sh /models/g2p/sequitur/althingi/g2p.mdl \
 temp/g2p_words.txt
derp	t ɛː r p
orð	ɔ r ð
stack usage:  200

Template

Here’s a template for adding new Icelandic g2p repos.

## repo link
#### Type
#### Location
#### Dependencies:
#### Usage
#### Example

input

output