File size: 2,350 Bytes
fc55559 | 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 | ---
license: apache-2.0
library_name: naptrap
language:
- en
pipeline_tag: feature-extraction
tags:
- grapheme-to-phoneme
- g2p
- keyword-spotting
- wake-word
- pytorch
- safetensors
---
# g2p_en decoder: keyword text features for naptrap
These are the weights of the [g2p_en](https://github.com/Kyubyong/g2p)
grapheme-to-phoneme model (Kyubyong Park & Jongseok Kim, Apache-2.0, release
2.1.0), repackaged as safetensors for
[naptrap](https://github.com/androiddrew/naptrap). The values are unchanged.
naptrap does **not** use this model to output phoneme strings. The keyword
models, such as
[`androiddrew/naptrap-phonmatchnet-en`](https://huggingface.co/androiddrew/naptrap-phonmatchnet-en),
were trained on the decoder's **hidden states**. This is the "embedding" path
from the PhonMatchNet reference code:
1. Normalize the keyword: lowercase it, strip accents, and keep `[a-z']` words.
Numbers must be spelled out.
2. Encode each word's characters with the grapheme GRU.
3. Decode greedily (up to 20 steps), keeping each emitted 256-D decoder state.
4. Join the words with the decoder's `<pad>` embedding as a separator.
The result is a `[positions, 256]` float32 matrix.
> Not affiliated with or endorsed by the g2p_en authors. See `LICENSE` and `NOTICE`.
## Usage
```python
from huggingface_hub import snapshot_download
from naptrap import G2PTextEncoder
encoder = G2PTextEncoder.from_pretrained(
snapshot_download("androiddrew/g2p-en-decoder-pytorch", revision="v1")
)
features = encoder("hey computer") # torch.float32 [positions, 256]
```
## Accuracy of the PyTorch port
We compared naptrap's PyTorch implementation with the original NumPy
`G2p.embedding()` on 545 keywords. They included Speech Commands words, wake
phrases, words with apostrophes and accents, 350 random CMUdict words and 150
random multi-word phrases. Every output length matched, and the maximum
absolute difference was **4.6e-6**.
## Limitations
The model was trained on CMUdict pronunciations of American English words.
Unusual names, brand names and non-English words get its best-guess
pronunciation, which may not match how people say them. naptrap does not use
the dictionary-lookup, homograph or number-expansion steps of the g2p_en
package.
`conversion/convert.py` reproduces `model.safetensors` from the g2p_en 2.1.0
wheel.
|