File size: 5,774 Bytes
6b43fdd 98a43a7 6b43fdd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 |
---
license: cc-by-nc-sa-4.0
language:
- ru
---
# Omogre
## Russian Accentuator and IPA Transcriptor
A library for [`Python 3`](https://www.python.org/). Automatic stress placement and [IPA transcription](https://en.wikipedia.org/wiki/International_Phonetic_Alphabet) for the Russian language.
## Dependencies
Installing the library will also install [`Pytorch`](https://pytorch.org/) and [`Numpy`](https://numpy.org/). Additionally, for model downloading, [`tqdm`](https://tqdm.github.io/) and [`requests`](https://pypi.org/project/requests/) will be installed.
## Installation
### Using GIT
```bash
pip install git+https://github.com/omogr/omogre.git
```
### Using pip
Download the code from [GitHub](https://github.com/omogr/omogre). In the directory containing [`setup.py`](https://github.com/omogr/omogre/blob/main/setup.py), run:
```bash
pip install -e .
```
### Manually
Download the code from [GitHub](https://github.com/omogr/omogre). Install [`Pytorch`](https://pytorch.org/), [`Numpy`](https://numpy.org/), [`tqdm`](https://tqdm.github.io/), and [`requests`](https://pypi.org/project/requests/). Run [`test.py`](https://github.com/omogr/omogre/blob/main/test.py).
## Model downloading
By default, data for models will be downloaded on the first run of the library. The script [`download_data.py`](https://github.com/omogr/omogre/blob/main/download_data.py) can also be used to download this data.
You can specify a path where the model data should be stored. If data already exists in this directory, it won't be downloaded again.
## Example
Script [`test.py`](https://github.com/omogr/omogre/blob/main/test.py).
```python
from omogre import Accentuator, Transcriptor
# Data will be downloaded to the 'omogre_data' directory
transcriptor = Transcriptor(data_path='omogre_data')
accentuator = Accentuator(data_path='omogre_data')
sentence_list = ['стены замка']
print('transcriptor', transcriptor(sentence_list))
print('accentuator', accentuator(sentence_list))
# Alternative call methods, differing only in notation
print('transcriptor.transcribe', transcriptor.transcribe(sentence_list))
print('accentuator.accentuate', accentuator.accentuate(sentence_list))
print('transcriptor.accentuate', transcriptor.accentuate(sentence_list))
```
## Class Parameters
### Transcriptor
All initialization parameters for the class are optional.
```python
class Transcriptor(data_path: str = None,
download: bool = True,
device_name: str = None,
punct: str = '.,!?')
```
- `data_path`: Directory where the model should be located.
- `device_name`: Parameter defining GPU usage. Corresponds to the initialization parameter of [torch.device](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device). Valid values include `"cpu"`, `"cuda"`, `"cuda:0"`, etc. Defaults to `"cuda"` if GPU is available, otherwise `"cpu"`.
- `punct`: List of non-letter characters to be carried over from the source text to the transcription. Default is `'.,!?'`.
- `download`: Whether to download the model from the internet if not found in `data_path`. Default is `True`.
Class methods:
```python
accentuate(sentence_list: list) -> list
transcribe(sentence_list: list) -> list
```
`accentuate` places stresses, `transcribe` performs transcription. Both inputs take a list of strings and return a list of strings.
### Accentuator
The `Accentuator` class for stress placement is identical to the `Transcriptor` in terms of stress functionality, except it doesn't load transcription data, reducing initialization time and memory usage.
All initialization parameters are optional, with the same meanings as for `Transcriptor`.
```python
class Accentuator(data_path: str = None,
download: bool = True,
device_name: str = None)
```
- `data_path`: Directory where the model should be located.
- `device_name`: Parameter for GPU usage. See above for details.
- `download`: Whether to download the model if not found. Default is `True`.
Class method:
```python
accentuate(sentence_list: list) -> list
```
## Usage Example
The script [`ruslan_markup.py`](https://github.com/omogr/omogre/blob/main/ruslan_markup.py) places stresses and generates transcriptions for markup files of the acoustic corpora [`ruslan`](http://dataset.sova.ai/SOVA-TTS/ruslan/ruslan_dataset.tar) and [`natasha`](http://dataset.sova.ai/SOVA-TTS/natasha/natasha_dataset.tar).
These markup files already contain manually placed stresses, which were [done manually](https://habr.com/ru/companies/ashmanov_net/articles/528296/).
The script [`ruslan_markup.py`](https://github.com/omogr/omogre/blob/main/ruslan_markup.py) generates its own stress placement for these files, allowing for an evaluation of the accuracy of stress placement.
## Context Awareness and Other Features
### Stresses
Stresses are placed considering context. If very long strings are encountered (for the current model, more than 510 tokens), context won't be considered for these. Stresses in these strings will be placed only where possible without context.
Stresses are also placed in one-syllable words, which might look unusual but simplifies subsequent transcription determination.
### Transcription
During transcription generation, extraneous characters are filtered out. Non-letter characters that are not filtered can be specified by a parameter. By default, four punctuation marks (`.,!?`) are not filtered. Transcription is determined word by word, without context. The following symbols are used for transcription:
```
ʲ`ɪətrsɐnjvmapkɨʊleɫdizofʂɕbɡxːuʐæɵʉɛ
```
## Feedback
Email for questions, comments and suggestions - `omogrus@ya.ru`.
---
license: cc-by-nc-sa-4.0
--- |