Update README.md
Browse files
README.md
CHANGED
@@ -26,6 +26,7 @@ ru_whisper_small is a fine-tuned version of [openai/whisper-small](https://huggi
|
|
26 |
|
27 |
## Intended uses & limitations
|
28 |
|
|
|
29 |
from transformers import WhisperProcessor, WhisperForConditionalGeneration
|
30 |
from datasets import load_dataset
|
31 |
|
@@ -45,12 +46,14 @@ predicted_ids = model.generate(input_features)
|
|
45 |
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=False)
|
46 |
|
47 |
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
|
|
|
48 |
|
49 |
|
50 |
## Long-Form Transcription
|
51 |
|
52 |
The Whisper model is intrinsically designed to work on audio samples of up to 30s in duration. However, by using a chunking algorithm, it can be used to transcribe audio samples of up to arbitrary length. This is possible through Transformers pipeline method. Chunking is enabled by setting chunk_length_s=30 when instantiating the pipeline. With chunking enabled, the pipeline can be run with batched inference. It can also be extended to predict sequence level timestamps by passing return_timestamps=True:
|
53 |
|
|
|
54 |
import torch
|
55 |
from transformers import pipeline
|
56 |
from datasets import load_dataset
|
@@ -71,12 +74,15 @@ prediction = pipe(sample.copy(), batch_size=8)["text"]
|
|
71 |
|
72 |
# we can also return timestamps for the predictions
|
73 |
prediction = pipe(sample.copy(), batch_size=8, return_timestamps=True)["chunks"]
|
|
|
74 |
|
75 |
|
76 |
## Faster using with Speculative Decoding
|
77 |
|
78 |
Speculative Decoding was proposed in Fast Inference from Transformers via Speculative Decoding by Yaniv Leviathan et. al. from Google. It works on the premise that a faster, assistant model very often generates the same tokens as a larger main model.
|
79 |
|
|
|
|
|
80 |
import torch
|
81 |
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
|
82 |
|
@@ -129,6 +135,7 @@ pipe = pipeline(
|
|
129 |
sample = dataset[0]["audio"]
|
130 |
result = pipe(sample)
|
131 |
print(result["text"])
|
|
|
132 |
|
133 |
|
134 |
### Training hyperparameters
|
|
|
26 |
|
27 |
## Intended uses & limitations
|
28 |
|
29 |
+
```bash
|
30 |
from transformers import WhisperProcessor, WhisperForConditionalGeneration
|
31 |
from datasets import load_dataset
|
32 |
|
|
|
46 |
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=False)
|
47 |
|
48 |
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
|
49 |
+
```
|
50 |
|
51 |
|
52 |
## Long-Form Transcription
|
53 |
|
54 |
The Whisper model is intrinsically designed to work on audio samples of up to 30s in duration. However, by using a chunking algorithm, it can be used to transcribe audio samples of up to arbitrary length. This is possible through Transformers pipeline method. Chunking is enabled by setting chunk_length_s=30 when instantiating the pipeline. With chunking enabled, the pipeline can be run with batched inference. It can also be extended to predict sequence level timestamps by passing return_timestamps=True:
|
55 |
|
56 |
+
```bash
|
57 |
import torch
|
58 |
from transformers import pipeline
|
59 |
from datasets import load_dataset
|
|
|
74 |
|
75 |
# we can also return timestamps for the predictions
|
76 |
prediction = pipe(sample.copy(), batch_size=8, return_timestamps=True)["chunks"]
|
77 |
+
```
|
78 |
|
79 |
|
80 |
## Faster using with Speculative Decoding
|
81 |
|
82 |
Speculative Decoding was proposed in Fast Inference from Transformers via Speculative Decoding by Yaniv Leviathan et. al. from Google. It works on the premise that a faster, assistant model very often generates the same tokens as a larger main model.
|
83 |
|
84 |
+
|
85 |
+
```bash
|
86 |
import torch
|
87 |
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
|
88 |
|
|
|
135 |
sample = dataset[0]["audio"]
|
136 |
result = pipe(sample)
|
137 |
print(result["text"])
|
138 |
+
```
|
139 |
|
140 |
|
141 |
### Training hyperparameters
|