Spaces:
Running
Running
File size: 9,810 Bytes
fe98679 874d788 fe98679 caed5a1 78740e3 fe98679 874d788 fe98679 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 |
<!DOCTYPE html>
<html>
<head>
<title>Hibiki</title>
<link
href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css"
rel="stylesheet"
/>
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/font/bootstrap-icons.min.css">
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.6.0/jquery.min.js"></script>
<script src="https://unpkg.com/wavesurfer.js@7"></script>
<script src="helper.js" defer></script>
<script>
function _setup_callback(elem, elems) {
elem.addEventListener("play", function () {
for (other of elems) {
if (other !== elem) {
other.pause();
}
}
});
}
document.addEventListener('DOMContentLoaded', function () {
var elems = document.body.getElementsByTagName("audio");
for (elem of elems) {
_setup_callback(elem, elems);
}
});
</script>
<style>
td {
vertical-align: middle;
text-align: center;
}
audio {
width: 20vw;
min-width: 100px;
max-width: 100%;
}
h1, h2, h3, h4, h5, h6, body, b, strong, th {
color: #595959;
}
.ratio-8x5 {
--bs-aspect-ratio: 62.5%;
}
.btn-secondary {
padding: 0.1rem 0.8rem;
font-size: small
}
.container {
max-width: 1620px;
}
</style>
</head>
<body>
<div class="container pt-5 mt-5 shadow p-5 mb-5 bg-white rounded">
<div class="text-center">
<h1>High-Fidelity Simultaneous Speech-To-Speech Translation</h1>
<p class="lead">
<a href="https://kyutai.org" target="_blank" rel="noopener noreferrer">Kyutai</a>
- code on <a href="https://github.com/kyutai-labs/hibiki" target="_blank" rel="noopener noreferrer">github</a>
</p>
</div>
<p>
<b>Abstract.</b>
We introduce <i>Hibiki</i> ('echo' in Japanese)
Hibiki leverages a multistream language model to synchronously process
source and target speech, and jointly produces text and audio tokens to
perform speech-to-text and speech-to-speech translation.
We furthermore address the fundamental challenge of <i>simultaneous</i> interpretation,
which unlike its <i>consecutive</i> counterpart---where one waits for
the end of the source utterance to start translating--- adapts its flow
to accumulate just enough context to produce a correct translation in
real-time, chunk by chunk. <br />
To do so, we introduce a weakly-supervised method that leverages the
perplexity of an off-the-shelf text translation system to identify
optimal delays on a per-word basis and create aligned synthetic data.
After supervised training, Hibiki performs adaptive, simultaneous
speech translation with vanilla temperature sampling. On a
French-English simultaneous speech translation task, Hibiki demonstrates
state-of-the-art performance in translation quality, speaker fidelity
and naturalness. Moreover, the simplicity of its inference process
makes it compatible with batched translation and even real-time
on-device deployment.
</p>
</div>
<div class="container shadow p-5 mb-5 bg-white rounded">
<h3>In the Wild Examples<a id="vis"/></h3>
<p class="mb-0">
</p>
<div class="container pt-3 table-responsive">
<table class="table table-hover" width="100%">
<tr>
<td witdth="50%">
<video class="embed-responsive-item" style="max-width: 80%; min-width: 400px;" controls>
<source src="videos/RPckvIkNWhE_ss301_to390_babel_numerique_arte.mp4" type="video/mp4">
Your browser does not support HTML video.
</video>
</td>
<td width="50%">
<video class="embed-responsive-item" style="max-width: 80%; min-width: 400px;" controls>
<source src="videos/uNAmODXvAiQ_ss9_message_a_caractere_informatif.mp4" type="video/mp4">
Your browser does not support HTML video.
</video>
</td>
<tr>
<td>
This example comes from a video explaining automated translation.
(<a href="https://www.youtube.com/watch?v=RPckvIkNWhE" target="_blank">source</a>, original video (c) Arte)
</td>
<td>
This example comes from a humoristic video. The source voice is high pitch on purpose,
it is a good showcase of how well Hibiki replicates pitch and prosody and how robust it is to
background noise <b>as no denoising is applied to the audio which is fed raw to Hibiki</b>.
(<a href="https://www.youtube.com/watch?v=uNAmODXvAiQ" target="_blank">source</a>, original video (c) Canal+)
</td>
</tr>
</table>
</div>
</div>
<div class="container shadow p-5 mb-5 bg-white rounded">
<h3>Examples with Ground Truth Interpretation<a id="vis"/></h3>
<p class="mb-0">
These samples come from the VoxPopuli dataset where the ground truth is real human
interpretation.
The volume for the sources has been reduced so that it's easier to hear the translations.
</p>
<div class="container pt-3 table-responsive">
<table class="table table-hover" id="vis-table2"></table>
</div>
</div>
<div class="container shadow p-5 mb-5 bg-white rounded">
<h3>Multistream Visualization<a id="vis"/></h3>
<p class="mb-0">
The audio for the source and translated versions are on different channels. Use headphones
to hear both at the same time. These samples are the same as in the voxpopuli section with CFG
set to 3.
</p>
<div class="container pt-3 table-responsive">
<table class="table table-hover" id="vis-table"></table>
</div>
</div>
<div class="container shadow p-5 mb-5 bg-white rounded">
<h3>Impact of Classifier-Free Guidance<a id="voxpopuli"/></h3>
<p class="mb-0">
Samples taken from the VoxPopuli dataset. The Hibiki samples are presented with different levels
of classifier-free guidance (CFG). The higher the CFG value, the closer the generated voice will
be to the original voice. This results in very strong accents for the generations with the higher
values.
</p>
<div class="container pt-3 table-responsive">
<table
class="table table-hover"
id="voxpopuli-table"
>
<thead>
<tr>
<th style="text-align: center">Source</th>
<th style="text-align: center">Hibiki CFG-1</th>
<th style="text-align: center">Hibiki CFG-3</th>
<th style="text-align: center">Hibiki CFG-10</th>
<th style="text-align: center">Seamless</th>
</tr>
</thead>
<tbody>
<tr> <td></td> <td></td> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td> <td></td> <td></td></tr>
</tbody>
</table>
</div>
</div>
<div class="container shadow p-5 mb-5 bg-white rounded">
<h3>Long-form Simultaneous Translations<a id="ntrex"/></h3>
<p class="mb-0">
Samples taken from the audio NTREX dataset.
</p>
<div class="container pt-3 table-responsive">
<table
class="table table-hover"
id="ntrex-table"
>
<thead>
<tr>
<th style="text-align: center;min-width: 200px;">Source</th>
<th style="text-align: center;">Hibiki</th>
<th style="text-align: center">Seamless</th>
</tr>
</thead>
<tbody>
<tr> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td></tr>
</tbody>
</table>
</div>
</div>
<div class="container shadow p-5 mb-5 bg-white rounded">
<h3>Short-form Simultaneous Translations<a id="cvss-c"/></h3>
<p class="mb-0">
Samples taken from the CVSS-C dataset.
</p>
<div class="container pt-3 table-responsive">
<table
class="table table-hover"
id="cvss-table"
>
<thead>
<tr>
<th style="text-align: center;min-width: 200px;">Source</th>
<th style="text-align: center;">Hibiki</th>
<th style="text-align: center">Seamless</th>
</tr>
</thead>
<tbody>
<tr> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td></tr>
<tr> <td></td> <td></td> <td></td></tr>
</tbody>
</table>
</div>
</div>
<div class="container p-5 mb-5 bg-white rounded">
<p class="mb-0">
This page was adapted from the <a href="https://google-research.github.io/seanet/soundstorm/examples">SoundStorm project page</a>.
</p>
</div>
</body>
</html>
|