Automatic Speech Recognition
Transformers
Safetensors
Russian
gigaam-ctc
asr
gigaam
stt
ctc
ngram
audio
speech
custom_code
Instructions to use waveletdeboshir/gigaam-ctc-with-lm with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use waveletdeboshir/gigaam-ctc-with-lm with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="waveletdeboshir/gigaam-ctc-with-lm", trust_remote_code=True)# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("waveletdeboshir/gigaam-ctc-with-lm", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
Update README.md
Browse files
README.md
CHANGED
|
@@ -14,7 +14,7 @@ tags:
|
|
| 14 |
- audio
|
| 15 |
- speech
|
| 16 |
---
|
| 17 |
-
|
| 18 |
|
| 19 |
# GigaAM-v2-CTC with ngram LM and beamsearch 🤗 Hugging Face transformers
|
| 20 |
|
|
@@ -66,8 +66,13 @@ input_features = processor(wav[0], sampling_rate=16000, return_tensors="pt")
|
|
| 66 |
with torch.no_grad():
|
| 67 |
logits = model(**input_features).logits
|
| 68 |
|
| 69 |
-
# decoding with beamseach and LM
|
| 70 |
-
transcription = processor.batch_decode(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
|
| 72 |
```
|
| 73 |
|
|
@@ -78,7 +83,13 @@ In our case (Conformer) `MODEL_STRIDE = 40` ms per timestamp.
|
|
| 78 |
|
| 79 |
```python
|
| 80 |
MODEL_STRIDE = 40
|
| 81 |
-
outputs = processor.batch_decode(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
word_ts = [
|
| 83 |
{
|
| 84 |
"word": d["word"],
|
|
|
|
| 14 |
- audio
|
| 15 |
- speech
|
| 16 |
---
|
| 17 |
+
[](https://colab.research.google.com/gist/waveletdeboshir/07e39ae96f27331aa3e1e053c2c2f9e8/gigaam-ctc-hf-with-lm.ipynb)
|
| 18 |
|
| 19 |
# GigaAM-v2-CTC with ngram LM and beamsearch 🤗 Hugging Face transformers
|
| 20 |
|
|
|
|
| 66 |
with torch.no_grad():
|
| 67 |
logits = model(**input_features).logits
|
| 68 |
|
| 69 |
+
# decoding with beamseach and LM (tune alpha, beta, beam_width for your data)
|
| 70 |
+
transcription = processor.batch_decode(
|
| 71 |
+
logits=logits.numpy(),
|
| 72 |
+
beam_width=64,
|
| 73 |
+
alpha=0.5,
|
| 74 |
+
beta=0.5,
|
| 75 |
+
).text[0]
|
| 76 |
|
| 77 |
```
|
| 78 |
|
|
|
|
| 83 |
|
| 84 |
```python
|
| 85 |
MODEL_STRIDE = 40
|
| 86 |
+
outputs = processor.batch_decode(
|
| 87 |
+
logits=logits.numpy(),
|
| 88 |
+
beam_width=64,
|
| 89 |
+
alpha=0.5,
|
| 90 |
+
beta=0.5,
|
| 91 |
+
output_word_offsets=True
|
| 92 |
+
)
|
| 93 |
word_ts = [
|
| 94 |
{
|
| 95 |
"word": d["word"],
|