Hugging Face's logo

Spaces:
Yehor
/
evaluate-asr-outputs
Running

App Files Community
evaluate-asr-outputs
File size: 5,019 Bytes 
5545d25
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5ab9924
5545d25
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5ab9924
5545d25
 
 
5ab9924
 
5545d25
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3727315
 
5545d25
 
 
 
 
 
5ab9924
 
 
 
 
 
 
 
 
 
5545d25
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5ab9924
 
 
5545d25
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5ab9924
5545d25
 
 
 
 
 
 
 
 
 
 
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
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
 
import sys

from importlib.metadata import version

import evaluate
import polars as pl
import gradio as gr

# Load evaluators
wer = evaluate.load("wer")
cer = evaluate.load("cer")

# Config
concurrency_limit = 5

title = "Evaluate ASR Outputs"

# https://www.tablesgenerator.com/markdown_tables
authors_table = """
## Authors

Follow them on social networks and **contact** if you need any help or have any questions:

| <img src="https://avatars.githubusercontent.com/u/7875085?v=4" width="100"> **Yehor Smoliakov** |
|-------------------------------------------------------------------------------------------------|
| https://t.me/smlkw in Telegram                                                                  |
| https://x.com/yehor_smoliakov at X                                                              |
| https://github.com/egorsmkv at GitHub                                                           |
| https://huggingface.co/Yehor at Hugging Face                                                    |
| or use [email protected]                                                                       |
""".strip()

examples = [
    ["evaluation_results.jsonl", True, False],
]

description_head = f"""
# {title}

## Overview

Upload a JSONL file generated by the ASR model.
""".strip()

description_foot = f"""
{authors_table}
""".strip()

metrics_value = """
Metrics will appear here.
""".strip()

tech_env = f"""
#### Environment

- Python: {sys.version}
""".strip()

tech_libraries = f"""
#### Libraries

- evaluate: {version('evaluate')}
- gradio: {version('gradio')}
- jiwer: {version('jiwer')}
- polars: {version('polars')}
""".strip()


def clean_value(x):
    return x.replace('’', "'").strip().lower().replace(',', '').replace('.', '').replace('?', '').replace('!', '').replace('–', '').replace('«', '').replace('»', '')

    
def inference(file_name, clear_punctuation, show_chars, progress=gr.Progress()):
    if not file_name:
        raise gr.Error("Please paste your JSON file.")

    progress(0, desc="Calculating...")

    df = pl.read_ndjson(file_name)

    inference_seconds = df['inference_total'].sum()
    duration_seconds = df['duration'].sum()

    rtf = inference_seconds / duration_seconds

    references = df['reference']

    if clear_punctuation:
        predictions = df['prediction'].map_elements(clean_value)
    else:
        predictions = df['prediction']

    # Evaluate
    wer_value = round(
        wer.compute(predictions=predictions, references=references), 4
    )
    cer_value = round(
        cer.compute(predictions=predictions, references=references), 4
    )

    inference_time = inference_seconds
    audio_duration = duration_seconds

    rtf = inference_time / audio_duration

    results = []

    results.append(f"Metrics using `evaluate` library:")
    results.append('')
    results.append(f"- WER: {wer_value} metric, {round(wer_value*100, 4)}%")
    results.append(f"- CER: {cer_value} metric, {round(cer_value*100, 4)}%")
    results.append('')
    results.append(f"- Accuracy on words: {round(100 - 100 * wer_value, 4)}%")
    results.append(f"- Accuracy on chars: {round(100 - 100 * cer_value, 4)}%")
    results.append('')
    results.append(f"- Inference time: {round(inference_time, 4)} seconds, {round(inference_time/60, 4)} mins, {round(inference_time/60/60, 4)} hours")
    results.append(f"- Audio duration: {round(audio_duration, 4)} seconds, {round(audio_duration/60/60, 4)} hours")
    results.append('')
    results.append(f"- RTF: {round(rtf, 4)}")

    if show_chars:
        all_chars = set()
        for pred in list(df['prediction']):
            for c in pred:
                all_chars.add(c)

        results.append('')
        results.append(f"Chars in predictions:")
        results.append(f"{list(all_chars)}")

    return "\n".join(results)


demo = gr.Blocks(
    title=title,
    analytics_enabled=False,
    theme=gr.themes.Base(),
)

with demo:
    gr.Markdown(description_head)

    gr.Markdown("## Usage")

    with gr.Row():
        with gr.Column():
            jsonl_file = gr.File(label="A JSONL file")
            clear_punctuation = gr.Checkbox(
                label="Clear punctuation, some chars and convert to lowercase",
            )
            show_chars = gr.Checkbox(
                label="Show chars in predictions",
            )

        metrics = gr.Textbox(
            label="Metrics",
            placeholder=metrics_value,
            show_copy_button=True,
        )

    gr.Button("Calculate").click(
        inference,
        concurrency_limit=concurrency_limit,
        inputs=[jsonl_file, clear_punctuation],
        outputs=metrics,
    )

    with gr.Row():
        gr.Examples(label="Choose an example", inputs=[jsonl_file, clear_punctuation, show_chars], examples=examples)

    gr.Markdown(description_foot)

    gr.Markdown("### Gradio app uses:")
    gr.Markdown(tech_env)
    gr.Markdown(tech_libraries)

if __name__ == "__main__":
    demo.queue()
    demo.launch()