Spaces:
Running
Running
File size: 20,220 Bytes
26600ec 8115b4a 98cede0 8115b4a 26600ec |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 |
<html><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta name="description" content="">
<meta name="author" content="">
<title> Continuous Audio Language Models </title>
<!-- Bootstrap core CSS -->
<link href="./page_files/bootstrap.min.css" rel="stylesheet">
<!-- Custom styles for this template -->
<link href="./page_files/scrolling-nav.css" rel="stylesheet">
<style>
.image-container {
display: flex;
justify-content: center;
align-items: center;
height: 50vh;
}
h2 {
text-decoration: underline;
}
.image-container img {
max-width: 75%;
max-height: 100%;
}
</style>
</head>
<body id="page-top">
<header class="bg-dark text-white">
<div class="container text-center">
<h1>Continuous Audio Language Models
</h1>
<p class="lead">
<a href="https://kyutai.org" target="_blank" rel="noopener noreferrer">Kyutai</a>
- paper on <a href="https://www.arxiv.org/abs/2509.06926" target="_blank" rel="noopener noreferrer">arxiv</a>
</p>
</div>
</header>
<div class="image-container">
<img src="src/figure.png" alt="Your Image">
</div>
<!-- Abstract Section -->
<section id="abstract" class="py-5">
<div class="container">
<h2>Abstract</h2>
<p>
Audio Language Models (ALM) have emerged as the dominant paradigm for speech and music generation by representing audio as sequences of discrete tokens. Yet, unlike text tokens, which are invertible, audio tokens are extracted from lossy codecs with a limited bitrate. As a consequence, increasing audio quality requires generating more tokens, which imposes a trade-off between fidelity and computational cost. We address this issue by studying Continuous Audio Language Models (CALM). These models instantiate a large Transformer backbone that produces a contextual embedding at every timestep. This sequential information then conditions an MLP that generates the next continuous frame of an audio VAE through consistency modeling. By avoiding lossy compression, CALM achieves higher quality at lower computational cost than their discrete counterpart. <br>
<br>
On this webpage, we show some results of our speech model as well as our music models. We illustrate as well the ablation study of the paper with some music samples.
</p>
</div>
</section>
<!-- Speech Language Model Section -->
<section id="speech" class="py-5">
<div class="container">
<h2>Speech Language Model</h2>
<p>
This section presents speech samples generated using a 3-second prompt. Key details of the setup and results include:
<ul>
<li>
Starting from a Helium pretrained 2B parameters text language model [1], we adopt the <strong>inner monologue framework</strong> from [2], where the model is fine-tuned with an audio stream delayed by 80 ms after the textual stream.</li>
</ul>
</li>
<li>
<strong>CALM setting:</strong> Audio stream is composed of continuous latents predicted via 1-step consistency modeling.
</li>
<li>
<strong>RQ-Transformer setting:</strong> Audio stream is produced using an 8-RVQ Mimi Codec and predicted in parallel by an RQ-Transformer.
</li>
<li>
<strong>Performance:</strong> CALM outperforms RQ-Transformer on meaningfulness. We believe this may be due to the backbone allocating less capacity to audio manipulation, leaving more for text prediction in the CALM setting. As well, we can see that temperature has a huge impact for both models, validating our heuristic for temperature sampling for CALM. </li>
<li>
<strong>Efficiency:</strong>
<ul>
<li>Sampling each latent from the consistency head is 12.3× faster than with the RQ-Transformer.</li>
<li>Generating 30 seconds of audio is overall 1.3× faster with CALM than with the baseline.</li>
</ul>
</li>
</ul>
</p>
<div class="table-responsive">
<table class="table table-bordered">
<thead class="thead-light">
<tr>
<th>Prompt</th>
<th>RQ-Transformer 8 RVQ<br>temp=0.8 (baseline)</th>
<th>CALM Consistency 1 Step<br>temp=0.8</th>
<th>CALM Consistency 1 Step<br>temp=1.0</th>
<th>RQ-Transformer 8 RVQ<br>temp=1.0</th>
</tr>
</thead>
<tbody>
<!-- Index 0 -->
<tr>
<td><audio controls src="src/speech/0/prompt.wav"></audio></td>
<td><audio controls src="src/speech/0/8rvq.wav"></audio></td>
<td><audio controls src="src/speech/0/consistency.wav"></audio></td>
<td><audio controls src="src/speech/0/consistency_no_temp.wav"></audio></td>
<td><audio controls src="src/speech/0/8rvq_no_temp.wav"></audio></td>
</tr>
<!-- Index 1 -->
<tr>
<td><audio controls src="src/speech/22/prompt.wav"></audio></td>
<td><audio controls src="src/speech/22/8rvq.wav"></audio></td>
<td><audio controls src="src/speech/22/consistency.wav"></audio></td>
<td><audio controls src="src/speech/22/consistency_no_temp.wav"></audio></td>
<td><audio controls src="src/speech/22/8rvq_no_temp.wav"></audio></td>
</tr>
<!-- Index 2 -->
<tr>
<td><audio controls src="src/speech/5/prompt.wav"></audio></td>
<td><audio controls src="src/speech/5/8rvq.wav"></audio></td>
<td><audio controls src="src/speech/5/consistency.wav"></audio></td>
<td><audio controls src="src/speech/5/consistency_no_temp.wav"></audio></td>
<td><audio controls src="src/speech/5/8rvq_no_temp.wav"></audio></td>
</tr>
<!-- Index 3 -->
<tr>
<td><audio controls src="src/speech/28/prompt.wav"></audio></td>
<td><audio controls src="src/speech/28/8rvq.wav"></audio></td>
<td><audio controls src="src/speech/28/consistency.wav"></audio></td>
<td><audio controls src="src/speech/28/consistency_no_temp.wav"></audio></td>
<td><audio controls src="src/speech/28/8rvq_no_temp.wav"></audio></td>
</tr>
<tr>
<td><audio controls src="src/speech/9/prompt.wav"></audio></td>
<td><audio controls src="src/speech/9/8rvq.wav"></audio></td>
<td><audio controls src="src/speech/9/consistency.wav"></audio></td>
<td><audio controls src="src/speech/9/consistency_no_temp.wav"></audio></td>
<td><audio controls src="src/speech/9/8rvq_no_temp.wav"></audio></td>
</tr>
<!-- Index 4 -->
<tr>
<td><audio controls src="src/speech/4/prompt.wav"></audio></td>
<td><audio controls src="src/speech/4/8rvq.wav"></audio></td>
<td><audio controls src="src/speech/4/consistency.wav"></audio></td>
<td><audio controls src="src/speech/4/consistency_no_temp.wav"></audio></td>
<td><audio controls src="src/speech/4/8rvq_no_temp.wav"></audio></td>
</tr>
</tbody>
</table>
</div>
</div>
</section>
<!-- Music Generation Section -->
<section id="music" class="py-5 bg-light">
<div class="container">
<h2>Music Generation</h2>
<p>
We compare our music generation models, all of which use a backbone with 1.35B parameters (from MusicGen Medium):
<ul>
<li><strong>Baseline:</strong> RQ-Transformer with 32 RVQ. </li>
<li><strong>CALM with TrigFlow (100 steps): </strong> Slower inference than the baseline.</li>
<li><strong>CALM with Consistency (4 steps): </strong> Inference is 1.9× faster than the baseline.</li>
<li><strong>CALM with Consistency (1 step): </strong> Inference is 2.1× faster than the baseline.</li>
<li><strong>Retrained MusicGen [3]:</strong> Trained on our dataset, with inference 1.3× faster than the baseline.</li>
</ul>
</p>
<div class="table-responsive">
<table class="table table-bordered">
<thead class="thead-light">
<tr>
<th>Prompt</th>
<th>RQ-Transformer 32 RVQ (baseline) FAD: 1.06</th>
<th>CALM TrigFlow 100 steps FAD: 0.64</th>
<th>CALM Consistency 4 steps FAD: 0.71</th>
<th>CALM Consistency 1 step FAD: 0.83</th>
<th>Retrained MusicGen FAD: 1.72</th>
</tr>
</thead>
<tbody>
<!-- Index 0 -->
<!-- Index 2 -->
<tr>
<td><audio controls src="src/music/2/prompt.wav"></audio></td>
<td><audio controls src="src/music/2/32rvq.wav"></audio></td>
<td><audio controls src="src/music/2/trigflow.wav"></audio></td>
<td><audio controls src="src/music/2/consistency_4.wav"></audio></td>
<td><audio controls src="src/music/2/consistency_1.wav"></audio></td>
<td><audio controls src="src/music/2/musicgen.wav"></audio></td>
</tr>
<!-- Index 3 -->
<tr>
<td><audio controls src="src/music/3/prompt.wav"></audio></td>
<td><audio controls src="src/music/3/32rvq.wav"></audio></td>
<td><audio controls src="src/music/3/trigflow.wav"></audio></td>
<td><audio controls src="src/music/3/consistency_4.wav"></audio></td>
<td><audio controls src="src/music/3/consistency_1.wav"></audio></td>
<td><audio controls src="src/music/3/musicgen.wav"></audio></td>
</tr>
<tr>
<td><audio controls src="src/music/9/prompt.wav"></audio></td>
<td><audio controls src="src/music/9/32rvq.wav"></audio></td>
<td><audio controls src="src/music/9/trigflow.wav"></audio></td>
<td><audio controls src="src/music/9/consistency_4.wav"></audio></td>
<td><audio controls src="src/music/9/consistency_1.wav"></audio></td>
<td><audio controls src="src/music/9/musicgen.wav"></audio></td>
</tr>
<tr>
<td><audio controls src="src/music/0/prompt.wav"></audio></td>
<td><audio controls src="src/music/0/32rvq.wav"></audio></td>
<td><audio controls src="src/music/0/trigflow.wav"></audio></td>
<td><audio controls src="src/music/0/consistency_4.wav"></audio></td>
<td><audio controls src="src/music/0/consistency_1.wav"></audio></td>
<td><audio controls src="src/music/0/musicgen.wav"></audio></td>
</tr>
<!-- Index 4 -->
<tr>
<td><audio controls src="src/music/4/prompt.wav"></audio></td>
<td><audio controls src="src/music/4/32rvq.wav"></audio></td>
<td><audio controls src="src/music/4/trigflow.wav"></audio></td>
<td><audio controls src="src/music/4/consistency_4.wav"></audio></td>
<td><audio controls src="src/music/4/consistency_1.wav"></audio></td>
<td><audio controls src="src/music/4/musicgen.wav"></audio></td>
</tr>
<!-- Index 4 -->
<!-- Index 4 -->
<tr>
<td><audio controls src="src/music/11/prompt.wav"></audio></td>
<td><audio controls src="src/music/11/32rvq.wav"></audio></td>
<td><audio controls src="src/music/11/trigflow.wav"></audio></td>
<td><audio controls src="src/music/11/consistency_4.wav"></audio></td>
<td><audio controls src="src/music/11/consistency_1.wav"></audio></td>
<td><audio controls src="src/music/11/musicgen.wav"></audio></td>
</tr>
<tr>
<td><audio controls src="src/music/6/prompt.wav"></audio></td>
<td><audio controls src="src/music/6/32rvq.wav"></audio></td>
<td><audio controls src="src/music/6/trigflow.wav"></audio></td>
<td><audio controls src="src/music/6/consistency_4.wav"></audio></td>
<td><audio controls src="src/music/6/consistency_1.wav"></audio></td>
<td><audio controls src="src/music/6/musicgen.wav"></audio></td>
</tr>
<!-- Index 4 -->
<tr>
<td><audio controls src="src/music/7/prompt.wav"></audio></td>
<td><audio controls src="src/music/7/32rvq.wav"></audio></td>
<td><audio controls src="src/music/7/trigflow.wav"></audio></td>
<td><audio controls src="src/music/7/consistency_4.wav"></audio></td>
<td><audio controls src="src/music/7/consistency_1.wav"></audio></td>
<td><audio controls src="src/music/7/musicgen.wav"></audio></td>
</tr>
</tbody>
</table>
</div>
</div>
</section>
<!-- Ablation Section -->
<section id="ablations" class="py-5 bg-light">
<div class="container">
<h2>Ablation Study</h2>
<p>
We illustrate here the ablation study of our paper in order to show the importance of each component of our model. We showcase it on Music Generation with CALM Consistency 4 steps. All the models have been trained 300k steps instead of 500k steps.
<ul>
<li><strong>Our model with Noise Augmentation, Short Context transformer and Head Batch Multiplier.</strong> CALM Consistency 4 steps.</li>
<li><strong>Without Noise Augmentation, Short Context transformer and Head Batch Multiplier.</strong> This configuration is very close to MAR [4]. It consists in just replacing the codec by a VAE and replacing the RQ-Transformer head by a consistency model. This model suffers from error accumulation, doesn't generate consistent music and tends to diverge (distorsion quickly happens).</li>
<li><strong>Without Short Context Transformer. </strong> Here, the addition of Noise Augmentation makes the model less sensitive to error accumulation. However, it quickly stops generating details focuses on the rhythmic and fades away during generation. This configuration is similar to the one used in [5] where the authors generate short single instrument music.</li>
<li><strong>Without Noise Augmentation. </strong> Here, there is no Noise Augmentation at training time but there is the Short Context Transformer that greatly helps the model to generate details. However there is still some error accumulation (see the 3rd sample).</li>
<li><strong>Without Head Batch Multiplier.</strong> Here, the model convergence is slower, its quality is lower for the same number of epochs.</li>
</ul>
</p>
<div class="table-responsive">
<table class="table table-bordered">
<thead class="thead-light">
<tr>
<th>Prompt</th>
<th>Our Model</th>
<th>Without Noise Aug., Short Context Transformer, Head Batch Mult. </th>
<th>Without Short Context Transformer</th>
<th>Without Noise Aug.</th>
<th>Without Head Batch Mult.</th>
</tr>
</thead>
<tbody>
<tr>
<td><audio controls src="src/ablations/0/prompt.wav"></audio></td>
<td><audio controls src="src/ablations/0/best_model.wav"></audio></td>
<td><audio controls src="src/ablations/0/no_noise_no_shortformer.wav"></audio></td>
<td><audio controls src="src/ablations/0/no_shortformer.wav"></audio></td>
<td><audio controls src="src/ablations/0/no_noise.wav"></audio></td>
<td><audio controls src="src/ablations/0/no_multiplier.wav"></audio></td>
</tr>
<tr>
<td><audio controls src="src/ablations/1/prompt.wav"></audio></td>
<td><audio controls src="src/ablations/1/best_model.wav"></audio></td>
<td><audio controls src="src/ablations/1/no_noise_no_shortformer.wav"></audio></td>
<td><audio controls src="src/ablations/1/no_shortformer.wav"></audio></td>
<td><audio controls src="src/ablations/1/no_noise.wav"></audio></td>
<td><audio controls src="src/ablations/1/no_multiplier.wav"></audio></td>
</tr>
<tr>
<td><audio controls src="src/ablations/2/prompt.wav"></audio></td>
<td><audio controls src="src/ablations/2/best_model.wav"></audio></td>
<td><audio controls src="src/ablations/2/no_noise_no_shortformer.wav"></audio></td>
<td><audio controls src="src/ablations/2/no_shortformer.wav"></audio></td>
<td><audio controls src="src/ablations/2/no_noise.wav"></audio></td>
<td><audio controls src="src/ablations/2/no_multiplier.wav"></audio></td>
</tr>
</tbody>
</table>
</div>
</div>
</section>
<!-- References Section -->
<section id="references" class="py-5 bg-light">
<div class="container">
<h2>References</h2>
<ul>
<li>
[1] Kyutai <em>"Helium 1: a modular and multilingual LLM"</em> <a href="https://kyutai.org/2025/04/30/helium.html" target="_blank">website</a>, 2025.
</li>
<li>
[2] Défossez, A., Mazaré, L., Orsini, M., Royer, A., Pérez, P., Jégou, H., Grave, E., & Zeghidour, N. (2024). <em>Moshi: A speech-text foundation model for real-time dialogue</em>. arXiv preprint arXiv:2410.00037. Available at: <a href="https://arxiv.org/abs/2410.00037" target="_blank">https://arxiv.org/abs/2410.00037</a>
</li>
<li>
[3] Copet, J., Kreuk, F., Gat, I., Remez, T., Kant, D., Synnaeve, G., Adi, Y., & Défossez, A. <em>"Simple and Controllable Music Generation."</em> <a href="https://arxiv.org/abs/2306.05284" target="_blank">arXiv:2306.05284</a>, 2023.
</li>
<li>
[4] Li, T., Tian, Y., Li, H., Deng, M., & He, K. (2024). <em>Autoregressive Image Generation Without Vector Quantization</em>. arXiv preprint arXiv:2406.11838. Available at: <a href="https://arxiv.org/abs/2406.11838" target="_blank">https://arxiv.org/abs/2406.11838</a>
</li>
<li>
[5] Pasini, M., Nistal, J., Lattner, S., & Fazekas, G. (2024). <em>Continuous Autoregressive Models with Noise Augmentation Avoid Error Accumulation</em>. arXiv preprint arXiv:2411.18447. Available at: <a href="https://arxiv.org/abs/2411.18447" target="_blank">https://arxiv.org/abs/2411.18447</a>
</li>
</ul>
</div>
</section>
<!-- Footer -->
<footer class="py-5 bg-dark">
<div class="container">
<!-- <p class="m-0 text-center text-white">Copyright © Your Website 2017</p> -->
</div>
<!-- /.container -->
</footer>
<!-- Bootstrap core JavaScript -->
<script src="./page_files/jquery.min.js"></script>
<script src="./page_files/bootstrap.bundle.min.js"></script>
<!-- Plugin JavaScript -->
<script src="./page_files/jquery.easing.min.js"></script>
<!-- Custom JavaScript for this theme -->
<script src="./page_files/scrolling-nav.js"></script>
<script> function setupCallback(elem, elems) {
elem.addEventListener("play", function () {
for (var other of elems) {
if (other !== elem) {
other.pause();
// other.currentTime = 0.;
}
}
});
}
document.addEventListener('DOMContentLoaded', function () {
var elems = document.body.getElementsByTagName("audio");
for (var elem of elems) {
setupCallback(elem, elems);
}
}); </script>
</body></html>
<script>
function setupCallback(elem, elems) {
elem.addEventListener("play", function () {
for (var other of elems) {
if (other !== elem) {
other.pause();
// other.currentTime = 0.;
}
}
});
}
document.addEventListener('DOMContentLoaded', function () {
var elems = document.body.getElementsByTagName("audio");
for (var elem of elems) {
setupCallback(elem, elems);
}
});
</script>
|