forked from llmware-ai/llmware
-
Notifications
You must be signed in to change notification settings - Fork 0
/
using-whisper-cpp-getting-started.py
86 lines (60 loc) · 3.5 KB
/
using-whisper-cpp-getting-started.py
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
""" This example shows how to get started using WhisperCPP as a fast, local voice-to-text processing engine.
Whisper is a leading open voice voice-to-text model from OpenAI - https://github.com/openai/whisper
WhisperCPP is the implementation of Whisper packaged as a GGML deliverable - https://github.com/ggerganov/whisper.cpp
Starting with llmware 0.2.11, we have integrated WhisperCPPModel as a new model class,
providing options for direct inference, and coming soon, integration into the Parser for easy text chunking and
parsing into a Library with other document types.
llmware provides prebuilt shared libraries for WhisperCPP on the following platforms:
--Mac M series
--Linux x86 (no CUDA)
--Linux x86 (with CUDA) - really fast
--Windows x86 (only on CPU) currently.
We have added three Whisper models to the default model catalog:
1. ggml-base.en.bin - english-only base model
2. ggml-base.bin - multi-lingual base model
2. ggml-small.en-tdrz.bin - this is a 'tiny-diarize' implementation that has been finetuned to identify the
speakers and inserts special [_SOLM_] tags to indicate a conversation turn / change of speaker.
Main repo: https://github.com/akashmjn/tinydiarize/
Citation: @software{mahajan2023tinydiarize,
author = {Mahajan, Akash}, month = {08},
title = {tinydiarize: Minimal extension of Whisper for speaker segmentation with special tokens},
url = {https://github.com/akashmjn/tinydiarize},
year = {2023}
To use WAV files, there is one additional Python dependency required:
--pip install librosa
--Note: this has been added to the default requirements.txt and pypy build starting with 0.2.11
To use other popular audio/video file formats, such as MP3, MP4, M4A, etc., then the following dependencies are
required:
--pip install pydub
--ffmpeg library - which can be installed as follows:
-- Linux: `sudo apt install ffmpeg'
-- Mac: `brew install ffmpeg`
-- Windows: direct download and install from ffmpeg
"""
import os
from llmware.models import ModelCatalog
from llmware.gguf_configs import GGUFConfigs
# optional / to adjust various log/display parameters of the model
GGUFConfigs().set_config("whisper_cpp_verbose", "OFF")
GGUFConfigs().set_config("whisper_cpp_realtime_display", True)
# note: english is default output - change to 'es' | 'fr' | 'de' | 'it' ...
GGUFConfigs().set_config("whisper_language", "en")
# whether to add or remove segment markers in llm response output
GGUFConfigs().set_config("whisper_remove_segment_markers", True)
def basic_whisper_cpp_use_example():
""" Hello world example to get started using WhisperCPP in LLMWare. """
fp = "/local/path/to/.wav"
fn = "my_wav.wav"
# prompt = string representing the path to a .wav file
prompt = os.path.join(fp,fn)
# choose between english-only and multilingual
whisper_base_english = "whisper-cpp-base-english"
whisper_base_multi = "whisper-cpp-base"
# load and run inference like any other model in llmware
model = ModelCatalog().load_model(whisper_base_english)
response = model.inference(prompt)
print("\nllm response: ", response["llm_response"])
print("usage: ", response["usage"])
return response
if __name__ == "__main__":
response = basic_whisper_cpp_use_example()