hf-speech-bench / app.py
sanchit-gandhi's picture
recommend latest CV
1c31846
raw
history blame
7.62 kB
import requests
import json
import pandas as pd
from tqdm.auto import tqdm
import streamlit as st
from huggingface_hub import HfApi, hf_hub_download
from huggingface_hub.repocard import metadata_load
aliases_lang = {"sv": "sv-SE"}
cer_langs = ["ja", "zh-CN", "zh-HK", "zh-TW"]
with open("languages.json") as f:
lang2name = json.load(f)
suggested_datasets = [
"librispeech_asr",
"mozilla-foundation/common_voice_8_0",
"mozilla-foundation/common_voice_11_0",
"speech-recognition-community-v2/eval_data",
"facebook/multilingual_librispeech"
]
def make_clickable(model_name):
link = "https://huggingface.co/" + model_name
return f'<a target="_blank" href="{link}">{model_name}</a>'
def get_model_ids():
api = HfApi()
models = api.list_models(filter="hf-asr-leaderboard")
model_ids = [x.modelId for x in models]
return model_ids
def get_metadata(model_id):
try:
readme_path = hf_hub_download(model_id, filename="README.md")
return metadata_load(readme_path)
except:
# 404 README.md not found
print(f"Model id: {model_id} is not great!")
return None
def parse_metric_value(value):
if isinstance(value, str):
"".join(value.split("%"))
try:
value = float(value)
except: # noqa: E722
value = None
elif isinstance(value, float) and value < 1.1:
# assuming that WER is given in 0.xx format
value = 100 * value
elif isinstance(value, list):
if len(value) > 0:
value = value[0]
else:
value = None
value = round(value, 2) if value is not None else None
return value
def parse_metrics_rows(meta):
if "model-index" not in meta or "language" not in meta:
return None
for result in meta["model-index"][0]["results"]:
if "dataset" not in result or "metrics" not in result:
continue
dataset = result["dataset"]["type"]
if "args" in result["dataset"] and "language" in result["dataset"]["args"]:
lang = result["dataset"]["args"]["language"]
else:
lang = meta["language"]
lang = lang[0] if isinstance(lang, list) else lang
lang = aliases_lang[lang] if lang in aliases_lang else lang
config = result["dataset"]["config"] if "config" in result["dataset"] else lang
split = result["dataset"]["split"] if "split" in result["dataset"] else None
row = {
"dataset": dataset,
"lang": lang,
"config": config,
"split": split
}
for metric in result["metrics"]:
type = metric["type"].lower().strip()
if type not in ["wer", "cer"]:
continue
value = parse_metric_value(metric["value"])
if value is None:
continue
if type not in row or value < row[type]:
# overwrite the metric if the new value is lower (e.g. with LM)
row[type] = value
if "wer" in row or "cer" in row:
yield row
@st.cache(ttl=600)
def get_data():
data = []
model_ids = get_model_ids()
for model_id in tqdm(model_ids):
meta = get_metadata(model_id)
if meta is None:
continue
for row in parse_metrics_rows(meta):
if row is None:
continue
row["model_id"] = model_id
data.append(row)
return pd.DataFrame.from_records(data)
def sort_datasets(datasets):
# 1. sort by name
datasets = sorted(datasets)
# 2. bring the suggested datasets to the top and append the rest
datasets = sorted(
datasets,
key=lambda dataset_id: suggested_datasets.index(dataset_id)
if dataset_id in suggested_datasets
else len(suggested_datasets),
)
return datasets
@st.cache(ttl=600)
def generate_dataset_info(datasets):
msg = """
The models have been trained and/or evaluated on the following datasets:
"""
for dataset_id in datasets:
if dataset_id in suggested_datasets:
msg += f"* [{dataset_id}](https://hf.co/datasets/{dataset_id}) *(recommended)*\n"
else:
msg += f"* [{dataset_id}](https://hf.co/datasets/{dataset_id})\n"
msg = "\n".join([line.strip() for line in msg.split("\n")])
return msg
dataframe = get_data()
dataframe = dataframe.fillna("")
st.sidebar.image("logo.png", width=200)
st.markdown("# The πŸ€— Speech Bench")
st.markdown(
f"This is a leaderboard of **{dataframe['model_id'].nunique()}** speech recognition models "
f"and **{dataframe['dataset'].nunique()}** datasets.\n\n"
"β¬… Please select the language you want to find a model for from the dropdown on the left."
)
lang = st.sidebar.selectbox(
"Language",
sorted(dataframe["lang"].unique(), key=lambda key: lang2name.get(key, key)),
format_func=lambda key: lang2name.get(key, key),
index=0,
)
lang_df = dataframe[dataframe.lang == lang]
sorted_datasets = sort_datasets(lang_df["dataset"].unique())
lang_name = lang2name[lang] if lang in lang2name else ""
num_models = len(lang_df["model_id"].unique())
num_datasets = len(lang_df["dataset"].unique())
text = f"""
For the `{lang}` ({lang_name}) language, there are currently `{num_models}` model(s)
trained on `{num_datasets}` dataset(s) available for `automatic-speech-recognition`.
"""
st.markdown(text)
st.sidebar.markdown("""
Choose the dataset that is most relevant to your task and select it from the dropdown below:
""")
dataset = st.sidebar.selectbox(
"Dataset",
sorted_datasets,
index=0,
)
dataset_df = lang_df[lang_df.dataset == dataset]
text = generate_dataset_info(sorted_datasets)
st.sidebar.markdown(text)
# sort by WER or CER depending on the language
metric_col = "cer" if lang in cer_langs else "wer"
if dataset_df["config"].nunique() > 1:
# if there are more than one dataset config
dataset_df = dataset_df[["model_id", "config", metric_col]]
dataset_df = dataset_df.pivot_table(index=['model_id'], columns=["config"], values=[metric_col])
dataset_df = dataset_df.reset_index(level=0)
else:
dataset_df = dataset_df[["model_id", metric_col]]
dataset_df.sort_values(dataset_df.columns[-1], inplace=True)
dataset_df = dataset_df.fillna("")
dataset_df.rename(
columns={
"model_id": "Model",
"wer": "WER (lower is better)",
"cer": "CER (lower is better)",
},
inplace=True,
)
st.markdown(
"Please click on the model's name to be redirected to its model card which includes documentation and examples on how to use it."
)
# display the model ranks
dataset_df = dataset_df.reset_index(drop=True)
dataset_df.index += 1
# turn the model ids into clickable links
dataset_df["Model"] = dataset_df["Model"].apply(make_clickable)
table_html = dataset_df.to_html(escape=False)
table_html = table_html.replace("<th>", '<th align="left">') # left-align the headers
st.write(table_html, unsafe_allow_html=True)
if lang in cer_langs:
st.markdown(
"---\n\* **CER** is [Char Error Rate](https://huggingface.co/metrics/cer)"
)
else:
st.markdown(
"---\n\* **WER** is [Word Error Rate](https://huggingface.co/metrics/wer)"
)
st.markdown(
"Want to beat the Leaderboard? Don't see your speech recognition model show up here? "
"Simply add the `hf-asr-leaderboard` tag to your model card alongside your evaluation metrics. "
"Try our [Metrics Editor](https://huggingface.co/spaces/huggingface/speech-bench-metrics-editor) to get started!"
)