rishidahiya/MiniMax-Speech-Tech-Report
0
1<!DOCTYPE html>2<html lang="en">3 4<head>5 <meta charset="UTF-8" />6 <meta name="viewport" content="width=device-width, initial-scale=1.0" />7 <title>MiniMax-Speech Tech Report | Intrinsic Zero-Shot Text-to-Speech with a Learnable Speaker Encoder</title>8 <meta name="description"9 content=" MiniMax-Speech, an autoregressive Transformer-based Text-to-Speech (TTS) model that generates high-quality speech" />10 <meta name="keywords" content="latex.css,css library,class-less css,latex css" />11 <meta property="og:title"12 content="MiniMax-Speech Tech Report | Intrinsic Zero-Shot Text-to-Speech with a Learnable Speaker Encoder" />13 <meta property="og:url" content="https://minimax-ai.github.io/tts_tech_report" />14 <meta property="og:description"15 content=" MiniMax-Speech, an autoregressive Transformer-based Text-to-Speech (TTS) model that generates high-quality speech" />16 <meta property="og:type" content="website" />17 18 <link rel="stylesheet" href="style.css" />19</head>20 21<body id="top" class="text-justify">22 <header23 style="background-image: url('assets/images/header-bg.jpeg'); background-size: cover; background-position: center; padding: 1rem 0; border-radius: 1rem;">24 <h1>MiniMax-Speech</h1>25 <h4 style="font-size: 1.3rem; line-height: 1; text-align: center;">Intrinsic Zero-Shot Text-to-Speech26 with a27 Learnable Speaker28 Encoder</h4>29 <p class="author">30 MiniMax Team <span class="date">May 2025</span><br />31 <a style="font-size: 1.1rem;" target="_blank" href="https://arxiv.org/abs/2505.07916">[Tech32 Report]</a>33 <a style="font-size: 1.1rem; margin-left: 1rem;" target="_blank"34 href="https://huggingface.co/datasets/MiniMaxAI/TTS-Multilingual-Test-Set">[Multilingual Test Set]</a>35 <a style="font-size: 1.1rem; margin-left: 1rem;" target="_blank" href="https://github.com/MiniMax-AI">[GitHub]</a>36 </p>37 </header>38 39 <div class="abstract">40 <h2>Abstract</h2>41 <p style="text-align: left;">42 We introduce MiniMax-Speech, an autoregressive Transformer-based Text-to-Speech (TTS) model that generates43 high-quality44 speech. A key innovation is our learnable speaker encoder, which extracts timbre features from a reference audio45 without46 requiring its transcription. This enables MiniMax-Speech to produce highly expressive speech with timbre47 consistent with48 the reference in a zero-shot manner, while also supporting one-shot voice cloning with exceptionally high49 similarity to50 the reference voice. In addition, the overall quality of the synthesized audio is enhanced through the proposed51 Flow-VAE. Our model supports 32 languages and demonstrates excellent performance across multiple objective and52 subjective evaluations metrics. Notably, it achieves state-of-the-art (SOTA) results on objective voice cloning53 metrics54 (Word Error Rate and Speaker Similarity) and has secured the top position on the public TTS Arena leaderboard.55 Another56 key strength of MiniMax-Speech, granted by the robust and disentangled representations from the speaker encoder,57 is its58 extensibility without modifying the base model, enabling various applications such as: arbitrary voice emotion59 control60 via LoRA; text to voice (T2V) by synthesizing timbre features directly from text description; and professional61 voice62 cloning (PVC) by fine-tuning timbre features with additional data.63 </p>64 </div>65 66 <nav role="navigation" class="toc">67 <h2>Explore MiniMax-Speech</h2>68 <p>Welcome to visit69 <a href="https://www.minimax.io/audio">MiniMax Audio</a> and70 explore our powerful TTS features.71 </p>72 <h2>Contents</h2>73 <ol>74 <li>75 <a href="#architecture-overview">Architecture Overview</a>76 </li>77 <li>78 <a href="#expressiveness-demonstrations">Expressiveness Demonstrations</a>79 <ol>80 <li><a href="#showcase-with-high-versatility">Showcase with High Versatility</a></li>81 <li><a href="#showcase-with-multiple-generation-attempts">Showcase with Multiple Generation Attempts</a></li>82 </ol>83 </li>84 <li><a href="#zero-shot-vs-one-shot-demonstrations">Zero-Shot vs. One-Shot Demonstrations</a></li>85 <li><a href="#multilingual-and-cross-lingual-capabilities-demonstrations">Multilingual and Cross-Lingual86 Capabilities Demonstrations</a></li>87 <li><a href="#flow-vae-vs-vae-comparisons">Flow-VAE vs. VAE Comparisons</a></li>88 <li><a href="#professional-voice-clone-pvc-demonstrations">Professional Voice Clone (PVC) Demonstrations</a></li>89 <li><a href="#emotion-control-demonstrations">Emotion Control Demonstrations</a></li>90 <li><a href="#text-prompted-voice-generation-demonstrations">Text-Prompted Voice Generation Demonstrations</a>91 </li>92 <li><a href="#comparison-of-voice-naturalness">Comparison of voice93 naturalness with the previous generation products</a></li>94 <li><a href="#citation">Citation</a></li>95 </ol>96 </nav>97 98 <main>99 <article>100 <div class="article-block">101 <h2 id="architecture-overview">Architecture Overview</h2>102 <figure>103 <img src="assets/images/system-overview.jpg" loading="lazy" alt="System Architecture" width="100%"104 height="auto" />105 <figcaption>106 An overview of the architecture of MiniMax-Speech.107 </figcaption>108 </figure>109 </div>110 111 <div class="article-block">112 <h2 id="expressiveness-demonstrations">Expressiveness Demonstrations</h2>113 <h3 id="showcase-with-high-versatility">Showcase with High Versatility</h3>114 <div class="scroll-wrapper">115 <table style="width: 100%;">116 <tbody>117 <tr class="border-bottom-thin">118 <th scope="col" style="width: 40%;">Description</th>119 <th scope="col" style="width: 30%; text-align: center;">Source Audio</th>120 <th scope="col" style="width: 30%; text-align: center;">Generated Audio</th>121 </tr>122 <tr class="border-bottom-thin">123 <td>124 A Compelling and Persuasive Speaker Voice125 </td>126 <td>127 <audio class="audio-md" src="assets/audios/Marketing_Voice_Sourse.wav" controls></audio>128 </td>129 <td>130 <audio class="audio-md" src="assets/audios/Compelling%20and%20Persuasive.wav" controls></audio>131 </td>132 </tr>133 <tr class="border-bottom-thin">134 <td>135 A Clear and Explanatory Voice with Broad Emotional Dynamics Across Different Texts136 </td>137 <td>138 <audio class="audio-md" src="assets/audios/Science_Voice_Sourse.wav" controls></audio>139 </td>140 <td>141 <audio class="audio-md" src="assets/audios/Explanatory%20Broad%20Emotional.wav" controls></audio>142 </td>143 </tr>144 <tr class="border-bottom-thin">145 <td>146 Another Explanatory Voice with Supernatural Prosody, <br>147 Featuring Distinct Ethnic and Age Characteristics148 </td>149 <td>150 <audio class="audio-md" src="assets/audios/Sociology_Sourse.mp3" controls></audio>151 </td>152 <td>153 <audio class="audio-md" src="assets/audios/Explanatory Supernatural Prosody.MP3" controls></audio>154 </td>155 </tr>156 <tr class="border-bottom-thin">157 <td>158 A Warm and Magnetic Voice that Brings Comfort159 </td>160 <td>161 <audio class="audio-md" src="assets/audios/Warm%20and%20Magnetic_Sourse.mp3" controls></audio>162 </td>163 <td>164 <audio class="audio-md" src="assets/audios/Warm%20and%20Magnetic.mp3" controls></audio>165 </td>166 </tr>167 <tr class="border-bottom-thin">168 <td>169 An ASMR Whispering Voice with Generated Breathing and Sound Effects170 </td>171 <td>172 <audio class="audio-md" src="assets/audios/Breathy%20ASMR_Sourse.wav" controls></audio>173 </td>174 <td>175 <audio class="audio-md" src="assets/audios/Breathy%20ASMR.MP3" controls></audio>176 </td>177 </tr>178 <tr class="border-bottom-thin">179 <td>180 A Robotic Voice with Rich Bass Resonance and Spatial Presence181 </td>182 <td>183 <audio class="audio-md" src="assets/audios/Lucky%20Robot_Sourse.wav" controls></audio>184 </td>185 <td>186 <audio class="audio-md" src="assets/audios/Lucky%20Robot.mp3" controls></audio>187 </td>188 </tr>189 <tr class="border-bottom-thin">190 <td>191 A Sardonic Mature Female Voice192 </td>193 <td>194 <audio class="audio-md" src="assets/audios/Onee-san_Sourse.MP3" controls></audio>195 </td>196 <td>197 <audio class="audio-md" src="assets/audios/Onee-san.wav" controls></audio>198 </td>199 </tr>200 </tbody>201 </table>202 </div>203 204 <h3 id="showcase-with-multiple-generation-attempts">Showcase with Multiple Generation Attempts, Post-Processing205 Audio Effects and Added Sound Effects</h3>206 <div class="scroll-wrapper">207 <table style="width: 100%;">208 <tbody>209 <tr class="border-bottom-thin">210 <th scope="col" style="width: 50%;">Description</th>211 <th scope="col" style="width: 50%; text-align: center;">Generated Audio</th>212 </tr>213 <tr class="border-bottom-thin">214 <td>215 A Husky Male Voice: From Soft Murmur to Excitement to Anger, then to Whispers216 </td>217 <td>218 <audio class="audio-lg" src="assets/audios/Murmur-Excitement-Anger-%20Whispers.MP3" controls></audio>219 </td>220 </tr>221 <tr class="border-bottom-thin">222 <td>223 An Angry Female Voice: From Soft Murmur to Rage to Reminiscence, then to Weeping224 </td>225 <td>226 <audio class="audio-lg" src="assets/audios/Neutral-Rage-Reminiscence-Weeping.MP3" controls></audio>227 </td>228 </tr>229 </tbody>230 </table>231 </div>232 </div>233 234 <div class="article-block">235 <h2 id="zero-shot-vs-one-shot-demonstrations">Zero-Shot vs. One-Shot Demonstrations</h2>236 <p>237 ZeroShot maintains speaker identity while generating more natural emotions, pauses, and other expressive238 features based239 on the text content, whereas OneShot adheres more strictly to the speaker characteristics (prosody, speech240 rate,241 emotions, etc.). For details of Zero-Shot and One-Shot, refer to the <a242 href="https://arxiv.org/abs/2505.07916" target="_blank">technical report</a>.243 </p>244 <div class="scroll-wrapper" style="margin-top: 2rem;">245 <table style="width: 100%;">246 <tbody>247 <tr class="border-bottom-thin">248 <th scope="col">Source Audio</th>249 <th scope="col">Text</th>250 <th scope="col">Zero-Shot Version</th>251 <th scope="col">One-Shot Version</th>252 <th scope="col">Elevenlabs Multilingual_v2</th>253 </tr>254 <tr class="border-bottom-thin">255 <td>256 <audio class="audio-sm" src="assets/audios/Lyrical%20Cantonese_Prompt.WAV" controls></audio>257 </td>258 <td>259 命运就算颠沛流离，<br>260 命运就算曲折离奇，<br>261 命运就算恐吓着你，<br>262 做人没趣味。<br>263 别流泪，心酸，更不应舍弃。<br>264 我愿能，一生永远陪伴你。265 </td>266 <td>267 <audio class="audio-sm" src="assets/audios/Lyrical%20Cantonese_ZeroShot.mp3" controls></audio>268 Preserving Distinctive Voice<br>269 Timbre and Expressive <br>270 Prosody with Regularized <br>271 Pausing and Speech Rate272 </td>273 <td>274 <audio class="audio-sm" src="assets/audios/Lyrical%20Cantonese_Oneshot.mp3" controls></audio>275 Better Reproduction of<br>276 Prompt's Exaggerated Speech<br>277 Rate and Characteristic<br>278 Phrase-Initial Pauses279 </td>280 <td>281 Cantonese not supported282 </td>283 </tr>284 <tr class="border-bottom-thin">285 <td>286 <audio class="audio-sm" src="assets/audios/Breaking%20Down%20Mandarin_Prompt.WAV" controls></audio>287 </td>288 <td>289 你们这些躲在道德高地的懦夫，<br>290 敢承认自己对本我的恐惧吗？<br>291 回答我！嗯？你回答我！<br>292 Look in my eyes！<br>293 老子写梦的解析时<br>294 你们还在玩泥巴，<br>295 我精神分析引论每个字母都能<br>296 刺穿文明社会的虚伪面具，<br>297 我解剖潜意识就像<br>298 外科医生划开皮肤。<br>299 是不是啊？说话！300 </td>301 <td>302 <audio class="audio-sm" src="assets/audios/Breaking%20Down%20Mandarin_ZeroShot.mp3" controls></audio>303 Capable of Generating<br>304 Relatively Calmer Emotions<br>305 while Preserving Voice<br>306 Identity307 </td>308 <td>309 <audio class="audio-sm" src="assets/audios/Breaking%20Down%20Mandarin_OneShot.mp3" controls></audio>310 Consistently Reproducing the<br>311 Angry Emotion from Prompt<br>312 in Every Utterance313 </td>314 <td>315 <audio class="audio-sm" src="assets/audios/ElevenLabs_Breaking Down Mandarin.mp3" controls></audio>316 </td>317 </tr>318 <tr class="border-bottom-thin">319 <td>320 <audio class="audio-sm" src="assets/audios/Quirky%20Female%20English_Prompt.MP3" controls></audio>321 </td>322 <td>323 Would you believe what happened at the<br>324 grocery store today? My goodness! The<br>325 avocados were on sale - half price! Half<br>326 price! I bought twenty of them!327 </td>328 <td>329 <audio class="audio-sm" src="assets/audios/Quirky%20Female%20English_ZeroShot.MP3" controls></audio>330 Effectively follows textual cues<br>331 for both longer and shorter<br>332 inter-sentence pauses333 </td>334 <td>335 <audio class="audio-sm" src="assets/audios/Quirky%20Female%20English_OneShot.MP3" controls></audio>336 Better reproduces the<br>337 exaggerated high pitch<br>338 characteristic of anime voices<br>339 </td>340 <td>341 <audio class="audio-sm" src="assets/audios/ElevenLabs_Quirky%20Female%20English.mp3" controls></audio>342 </td>343 </tr>344 <tr>345 <td>346 <audio class="audio-sm" src="assets/audios/Neurotic%20Teenage%20English_Prompt.MP3" controls></audio>347 </td>348 <td>349 Oh my gosh, like, I literally can't believe<br>350 what just happened! Um, so basically, I was,<br>351 you know, just sitting there in class,<br>352 right? And then, ugh, this totally weird<br>353 thing happened - like, seriously weird! Wait,<br>354 wait... Should I even be talking about this?<br>355 Ugh, whatever.356 </td>357 <td>358 <audio class="audio-sm" src="assets/audios/Neurotic%20Teenage%20English_ZeroShot.MP3"359 controls></audio>360 Effectively follows textual cues<br>361 for both longer and shorter<br>362 inter-sentence pauses363 </td>364 <td>365 <audio class="audio-sm" src="assets/audios/Neurotic%20Teenage%20English_OneShot.MP3" controls></audio>366 Better reproduces the<br>367 exaggerated high pitch<br>368 characteristic of anime voices<br>369 </td>370 <td>371 <audio class="audio-sm" src="assets/audios/ElevenLabs_Neurotic%20Teenage%20English.mp3"372 controls></audio>373 </td>374 </tr>375 </tbody>376 </table>377 </div>378 </div>379 380 <div class="article-block">381 <h2 id="multilingual-and-cross-lingual-capabilities-demonstrations">Multilingual and Cross-Lingual Capabilities382 Demonstrations</h2>383 <p>Speech-02-HD maintains high naturalness in less common languages while demonstrating significant advantages384 in385 Standard386 Chinese pronunciation accuracy.</p>387 <div class="scroll-wrapper" style="margin-top: 2rem;">388 <table style="width: 100%;">389 <tbody>390 <tr class="border-bottom-thin">391 <th scope="col">Languages</th>392 <th scope="col">Source Audio</th>393 <th scope="col">Text</th>394 <th scope="col">MiniMax<br>Speech_02_HD</th>395 <th scope="col">ElevenLabs<br>Multilingual_v2</th>396 <th scope="col">OpenAI<br>TTS_1_HD<br>(*not cloned voice)</th>397 </tr>398 <!-- Thai -->399 <tr class="border-bottom-thin">400 <th>Thai</th>401 <td>402 <audio class="audio-sm" src="assets/audios/Thai_Male_Sourse.wav" controls></audio>403 </td>404 <td>405 สวัสดีค่ะ วันนี้อากาศดีมากเลย<br>406 คุณจะไปทานอาหารกลางวันที่ไหนคะ<br>407 ฉันกำลังคิดว่าจะไปร้านอาหารไทยแถวนี้<br>408 </td>409 <td>410 <audio class="audio-sm" src="assets/audios/Thai.mp3" controls></audio>411 </td>412 <td>413 Thai not perfectly supported414 <audio class="audio-sm" src="assets/audios/ElevenLabs_Thai.mp3" controls></audio>415 </td>416 <td>417 <audio class="audio-sm" src="assets/audios/OpenAI_Thai.mp3" controls></audio>418 </td>419 </tr>420 <!-- Vietnamese -->421 <tr class="border-bottom-thin">422 <th>Vietnamese</th>423 <td>424 <audio class="audio-sm" src="assets/audios/Vietnamese_Female_Sourse.wav" controls></audio>425 </td>426 <td>427 Tôi đang đọc một cuốn sách rất hay về lịch sử Việt Nam.<br>428 Những câu chuyện về văn hóa truyền<br>429 thống thật sự rất thú vị.<br>430 </td>431 <td>432 <audio class="audio-sm" src="assets/audios/Vietnamese.mp3" controls></audio>433 </td>434 <td>435 Vietnamese not perfectly supported436 <audio class="audio-sm" src="assets/audios/ElevenLabs_Vietnamese.mp3" controls></audio>437 </td>438 <td>439 <audio class="audio-sm" src="assets/audios/OpenAI_Vietnamese.mp3" controls></audio>440 </td>441 </tr>442 <!-- Czech -->443 <tr class="border-bottom-thin">444 <th>Czech</th>445 <td>446 <audio class="audio-sm" src="assets/audios/Czech_Female_Sourse.wav" controls></audio>447 </td>448 <td>449 Ranní mlha se pomalu zvedá nad řekou,<br>450 zatímco první paprsky slunce prosvítají mezi stromy.<br>451 Ptáci začínají svůj ranní koncert.<br>452 </td>453 <td>454 <audio class="audio-sm" src="assets/audios/Czech.mp3" controls></audio>455 </td>456 <td>457 <audio class="audio-sm" src="assets/audios/ElevenLabs_Czech.mp3" controls></audio>458 </td>459 <td>460 <audio class="audio-sm" src="assets/audios/OpenAI_Czech.mp3" controls></audio>461 </td>462 </tr>463 <!-- Polish -->464 <tr class="border-bottom-thin">465 <th>Polish</th>466 <td>467 <audio class="audio-sm" src="assets/audios/Polish_Male_Sourse.wav" controls></audio>、468 </td>469 <td>470 Młoda sowa siedzi cicho na gałęzi sosny,<br>471 obserwując leśną polanę w świetle księżyca.<br>472 Wiatr delikatnie porusza liśćmi drzew.<br>473 </td>474 <td>475 <audio class="audio-sm" src="assets/audios/Polish.mp3" controls></audio>476 </td>477 <td>478 <audio class="audio-sm" src="assets/audios/ElevenLabs_Polish.mp3" controls></audio>479 </td>480 <td>481 <audio class="audio-sm" src="assets/audios/OpenAI_Polish.mp3" controls></audio>482 </td>483 </tr>484 <!-- Japanese -->485 <tr class="border-bottom-thin">486 <th>Japanese</th>487 <td>488 <audio class="audio-sm" src="assets/audios/Japanese_DominantMan_Sourse.mp3" controls></audio>489 </td>490 <td>491 電車が遅延している影響で、渋谷駅がとても混雑<br>492 しています。次の山手線は約10分後に到着<br>493 予定です。お急ぎのお客様は、他の路線も<br>494 ご利用ください。495 </td>496 <td>497 <audio class="audio-sm" src="assets/audios/Japanese.mp3" controls></audio>498 </td>499 <td>500 <audio class="audio-sm" src="assets/audios/ElevenLabs_Japanese_Dominant_Man.mp3" controls></audio>501 </td>502 <td>503 <audio class="audio-sm" src="assets/audios/OpenAI_Japanese.mp3" controls></audio>504 </td>505 </tr>506 </tbody>507 </table>508 </div>509 <p style="margin-top: 4rem;">Speech-02-HD has superior performance in zero-shot cross-lingual scenarios.</p>510 <div class="scroll-wrapper" style="margin-top: 2rem;">511 <table style="width: 100%;">512 <tbody>513 <tr class="border-bottom-thin">514 <th scope="col">Original Language</th>515 <th scope="col">Source Audio</th>516 <th scope="col">Mixed Language</th>517 <th scope="col">Text</th>518 <th scope="col">MiniMax<br>Speech_02_HD</th>519 <th scope="col">ElevenLabs<br>Multilingual_v2</th>520 <th scope="col">OpenAI<br>TTS_1_HD<br>(*not cloned voice)</th>521 </tr>522 <tr class="border-bottom-thin">523 <td>English</td>524 <td>525 <audio class="audio-sm" src="assets/audios/Wong_Sourse.mp3" controls></audio>526 </td>527 <td>English + Mandarin</td>528 <td>529 Kiddo! Come come come, 学如逆水行舟，不进则退。<br>530 I see you're using AI tools already - so smart!<br>531 But eh, cannot just rely on tools only lah!<br>532 The future belongs to those who can work alongside AI,<br>533 not those scared of it.534 </td>535 <td>536 <audio class="audio-md" src="assets/audios/English-Mandarin.wav" controls></audio>537 </td>538 <td>539 <audio class="audio-md" src="assets/audios/ElevenLabs_English-Mandarin.mp3" controls></audio>540 </td>541 <td>542 <audio class="audio-md" src="assets/audios/OpenAI_English-Mandarin.mp3" controls></audio>543 </td>544 </tr>545 <tr class="border-bottom-thin">546 <td>Mandarin</td>547 <td>548 <audio class="audio-sm" src="assets/audios/ShiBanYu_Sourse.mp3" controls></audio>549 </td>550 <td>Mandarin + Cantonese</td>551 <td>552 老铁啊，多谢晒你送我呢本，广州话正音字典，咁好嘢喎！<br>553 我呢个大老爷们儿学广州话真系好难㗎！成日都分唔清声调啊。<br>554 嗱，而家有咗呢本书，什么都好啦。555 </td>556 <td>557 <audio class="audio-md" src="assets/audios/Mandarin-Cantonese.MP3" controls></audio>558 </td>559 <td>560 Cantonese not supported561 </td>562 <td>563 Cantonese not supported564 </td>565 </tr>566 <tr class="border-bottom-thin">567 <td>Mandarin</td>568 <td>569 <audio class="audio-sm" src="assets/audios/ShuanQ_Sourse.mp3" controls></audio>570 </td>571 <td>Mandarin + English</td>572 <td>573 The people said, 桂林's scenery is the first under heaven.<br>574 Yet in my opinion, 阳朔 scenery is better than 桂林。<br>575 群峰倒影山浮水，无水无山不入神。576 </td>577 <td>578 <audio class="audio-md" src="assets/audios/Mandarin-English.WAV" controls></audio>579 </td>580 <td>581 <audio class="audio-md" src="assets/audios/ElevenLabs_Mandarin-English.mp3" controls></audio>582 </td>583 <td>584 <audio class="audio-md" src="assets/audios/OpenAI_Mandarin-English.mp3" controls></audio>585 </td>586 </tr>587 <tr class="border-bottom-thin">588 <td>English</td>589 <td>590 <audio class="audio-sm" src="assets/audios/CoCo_Sourse.mp3" controls></audio>591 </td>592 <td>English + Spanish</td>593 <td>594 Mi abuelita always told me "el que persevera, alcanza".<br>595 If you persevere, you'll achieve your dreams!<br>596 Guess what! They choose me to play the lead role in our BIG show!597 </td>598 <td>599 <audio class="audio-md" src="assets/audios/English-Spanish.wav" controls></audio>600 </td>601 <td>602 <audio class="audio-md" src="assets/audios/ElevenLabs_English-Spanish.mp3" controls></audio>603 </td>604 <td>605 <audio class="audio-md" src="assets/audios/OpenAI_English-Spanish.mp3" controls></audio>606 </td>607 </tr>608 <tr class="border-bottom-thin">609 <td>Japanese</td>610 <td>611 <audio class="audio-sm" src="assets/audios/Powerful_Girl_Sourse.mp3" controls></audio>612 </td>613 <td>Japanese + Korean</td>614 <td>615 最近の天気予報によりますと、今週末は桜の開花に最適<br>616 な気温になる予定です。<br>617 東京都内の各公園では花見客で賑わうことが予想されますが、<br>618 서울에서도 벚꽃이 피기 시작했다고 하네요.<br>619 이번 주말에는 여의도 공원에서 벚꽃 축제가 열린다고 하니<br>620 많은 분들이 찾아오실 것 같습니다.621 </td>622 <td>623 <audio class="audio-md" src="assets/audios/Japanese-Korean.mp3" controls></audio>624 </td>625 <td>626 <audio class="audio-md" src="assets/audios/ElevenLabs_Japanese-Korean.mp3" controls></audio>627 </td>628 <td>629 <audio class="audio-md" src="assets/audios/OpenAI_Japanese-Korean.mp3" controls></audio>630 </td>631 </tr>632 </tbody>633 </table>634 </div>635 <p>*Although OpenAI currently does not support voice cloning functionality, we still wish to conduct comparative636 listening637 tests with its excellent naturalness as a reference.</p>638 </div>639 640 <div class="article-block">641 <h2 id="flow-vae-vs-vae-comparisons">Flow-VAE vs. VAE Comparison</h2>642 <p>Flow-VAE is less likely to produce the following instabilities.</p>643 <div class="scroll-wrapper" style="margin-top: 2rem;">644 <table style="width: 100%;">645 <tbody>646 <tr class="border-bottom-thin">647 <th scope="col" style="text-align: center;">Source Audio</th>648 <th scope="col" style="text-align: center;">Flow-VAE</th>649 <th scope="col" style="text-align: center;">VAE</th>650 <th scope="col" style="text-align: center;">Differences</th>651 </tr>652 <tr class="border-bottom-thin">653 <td style="width: 25%">654 <audio src="assets/audios/Condition1.wav" controls></audio>655 </td>656 <td style="width: 25%">657 <audio src="assets/audios/FlowVAE1.wav" controls></audio>658 </td>659 <td style="width: 25%">660 <audio src="assets/audios/VAE1.wav" controls></audio>661 </td>662 <td>663 Flow-VAE reproduces more continuous<br>664 and natural reverberation665 </td>666 </tr>667 <tr class="border-bottom-thin">668 <td>669 <audio src="assets/audios/Condition2.wav" controls></audio>670 </td>671 <td>672 <audio src="assets/audios/FlowVAE2.wav" controls></audio>673 </td>674 <td>675 <audio src="assets/audios/VAE2.wav" controls></audio>676 </td>677 <td>678 VAE introduces unwanted<br>679 high-frequency components680 </td>681 </tr>682 <tr>683 <td>684 <audio src="assets/audios/Conditon3.wav" controls></audio>685 </td>686 <td>687 <audio src="assets/audios/FlowVAE3.wav" controls></audio>688 </td>689 <td>690 <audio src="assets/audios/VAE3.wav" controls></audio>691 </td>692 <td>693 VAE produces electronic-sounding<br>694 artifacts at the beginning695 </td>696 </tr>697 </tbody>698 </table>699 </div>700 </div>701 702 <div class="article-block">703 <h2 id="professional-voice-clone-pvc-demonstrations">Professional Voice Clone (PVC) Demonstrations</h2>704 <p>For more complex dialectal accents and tonal characteristics, PVC can reproduce these features while705 maintaining high706 naturalness based on the text content.</p>707 <div class="scroll-wrapper" style="margin-top: 2rem;">708 <table style="width: 100%;">709 <tbody>710 <tr class="border-bottom-thin">711 <th scope="col" style="text-align: center;">Source Audio</th>712 <th scope="col" style="text-align: center;">Zero-Shot</th>713 <th scope="col" style="text-align: center;">PVC</th>714 <th scope="col" style="text-align: center;">Differences</th>715 </tr>716 <tr class="border-bottom-thin">717 <td style="width: 25%">718 <audio src="assets/audios/JosephBrodsky_Source.wav" controls></audio>719 </td>720 <td style="width: 25%">721 <audio src="assets/audios/JosephBrodsky_Fast.mp3" controls></audio>722 </td>723 <td style="width: 25%">724 <audio src="assets/audios/JosephBrodsky_PVC.mp3" controls></audio>725 </td>726 <td>727 Like the ZeroShot version, the PVC<br>728 version has rising sentence-final intonation,<br>729 but distinctively sustains this<br>730 elevated pitch instead of the typical<br>731 pitch declination found in common<br>732 declarative sentences733 </td>734 </tr>735 <tr class="border-bottom-thin">736 <td>737 <audio src="assets/audios/TianJin_Source.wav" controls></audio>738 </td>739 <td>740 <audio src="assets/audios/TianJin_Fast.mp3" controls></audio>741 </td>742 <td>743 <audio src="assets/audios/TianJin_PVC.mp3" controls></audio>744 </td>745 <td>746 With more materials, the model not only<br>747 reproduces the speaker's voice characteristics<br>748 but also accurately captures more<br>749 dialectal features750 </td>751 </tr>752 </tbody>753 </table>754 </div>755 </div>756 757 <div class="article-block">758 <h2 id="emotion-control-demonstrations">Emotion Control Demonstrations</h2>759 <h3>Source Audio for Refreshing Young Man</h3>760 <audio src="assets/audios/Mandarin_Refreshing_Young_Man_Sourse.mp3" controls></audio>761 <h3>DEMO</h3>762 <div class="scroll-wrapper">763 <table style="width: 100%;">764 <tbody>765 <tr class="border-bottom-thin">766 <th scope="col">Neutral</th>767 <th scope="col" style="min-width: 120px;">Emotion</th>768 <th scope="col">Text</th>769 <th scope="col">Emotion Control Audio</th>770 </tr>771 <tr class="border-bottom-thin">772 <td>773 <audio class="audio-md" src="assets/audios/Neutral1.mp3" controls></audio>774 </td>775 <td>776 Surprised777 </td>778 <td>779 天哪！我完全没想到会在这里遇见你，<br>780 都过去这么多年了，你一点都没变！781 </td>782 <td>783 <audio class="audio-md" src="assets/audios/Surprised.mp3" controls></audio>784 </td>785 </tr>786 <tr class="border-bottom-thin">787 <td>788 <audio class="audio-md" src="assets/audios/Neutral2.mp3" controls></audio>789 </td>790 <td>791 Disgusted792 </td>793 <td>794 这个地方实在太脏乱了，到处都是垃圾和难闻的气味儿，<br>795 我一秒钟都不想多待。796 </td>797 <td>798 <audio class="audio-md" src="assets/audios/Disgusted.mp3" controls></audio>799 </td>800 </tr>801 <tr class="border-bottom-thin">802 <td>803 <audio class="audio-md" src="assets/audios/Neutral3.mp3" controls></audio>804 </td>805 <td>806 Fearful807 </td>808 <td>809 深夜回家的路上，我清楚地听见身后有脚步声在跟着我，<br>810 可是回头却什么都看不见。811 </td>812 <td>813 <audio class="audio-md" src="assets/audios/Fearful.mp3" controls></audio>814 </td>815 </tr>816 <tr class="border-bottom-thin">817 <td>818 <audio class="audio-md" src="assets/audios/Neutral4.mp3" controls></audio>819 </td>820 <td>821 Angry822 </td>823 <td>824 我付出了这么多，换来的却是这样的背叛！<br>825 你怎么可以这样对待我的信任！826 </td>827 <td>828 <audio class="audio-md" src="assets/audios/Angry.mp3" controls></audio>829 </td>830 </tr>831 <tr class="border-bottom-thin">832 <td>833 <audio class="audio-md" src="assets/audios/Neutral5.mp3" controls></audio>834 </td>835 <td>836 Sad837 </td>838 <td>839 躺在床上翻来覆去，心里压着说不出的难过和沮丧，<br>840 昨天晚上又失眠了。841 </td>842 <td>843 <audio class="audio-md" src="assets/audios/Sad.mp3" controls></audio>844 </td>845 </tr>846 <tr class="border-bottom-thin">847 <td>848 <audio class="audio-md" src="assets/audios/Neutral6.mp3" controls></audio>849 </td>850 <td>851 Happy852 </td>853 <td>854 和好朋友一起在院子里烧烤，聊着有趣的故事，<br>855 享受着美食和欢乐的时光。856 </td>857 <td>858 <audio class="audio-md" src="assets/audios/Happy.mp3" controls></audio>859 </td>860 </tr>861 </tbody>862 </table>863 </div>864 </div>865 866 <div class="article-block">867 <h2 id="text-prompted-voice-generation-demonstrations">Text-Prompted Voice Generation Demonstrations</h2>868 <div class="scroll-wrapper">869 <table style="width: 100%;">870 <tbody>871 <tr class="border-bottom-thin">872 <th scope="col">Prompt</th>873 <th scope="col">Text</th>874 <th scope="col" style="text-align: center;">Audio</th>875 </tr>876 <tr class="border-bottom-thin">877 <td>878 男性中年声音，说中文，音色浑厚醇厚，带有自然的磁性，<br>879 语速偏慢，音量适中，音调偏低沉。声音整体给人沉稳可靠的感觉，<br>880 在深度访谈场景中表现出专业性和亲和力，音质清晰，吐字规整有力。881 </td>882 <td>883 在这个安静的夜晚，让我们一起走进《人生笔记》这本书。<br>884 作者用平实的文字记录下生活中的点点滴滴，<br>885 让我们看到平凡中的真善美。<br>886 今天，我们先来读第一章：'生活的痕迹'......887 </td>888 <td>889 <audio class="audio-md" src="assets/audios/深度访谈男中年.wav" controls></audio>890 </td>891 </tr>892 <tr class="border-bottom-thin">893 <td>894 说中文的女青年，音色偏甜美，语速比较快，<br>895 说话时带着一种轻快的感觉，整体音调较高，像是在直播带货，<br>896 整体氛围比较活跃，声音清晰，听起来很有亲和力。897 </td>898 <td>899 亲爱的宝宝们，等了好久的神仙面霜终于到货啦！<br>900 你们看这个包装是不是超级精致？<br>901 我自己已经用了一个月了，效果真的绝绝子！<br>902 而且这次活动价真的太划算了，错过真的会后悔的哦～903 </td>904 <td>905 <audio class="audio-md" src="assets/audios/直播带货女青年.wav" controls></audio>906 </td>907 </tr>908 <tr class="border-bottom-thin">909 <td>910 中国男性声音，听着像是青年，音色清亮，语速比较快，<br>911 说话很有激情，像是在解说比赛，声音中带着紧张和兴奋的感觉。912 </td>913 <td>914 漂亮！这个进攻太精彩了！张伟突破防线，<br>915 一个漂亮的转身，球传到禁区，王超跟上，射门！<br>916 球进了！难以置信的精彩配合，现场观众都沸腾了！917 </td>918 <td>919 <audio class="audio-md" src="assets/audios/体育解说男青年.wav" controls></audio>920 </td>921 </tr>922 <tr class="border-bottom-thin">923 <td>924 中国女青年的声音，音色清脆，说话速度偏快，语调活泼，<br>925 像是在做游戏直播，声音中带着愉快的感觉，整体音调较高，<br>926 整体氛围比较轻松。927 </td>928 <td>929 啊！这里有个宝箱！让我们看看里面是什么～<br>930 哇！是传说中的紫色装备！运气也太好了吧！<br>931 谢谢小伙伴们的打赏，我们继续往前探索......932 </td>933 <td>934 <audio class="audio-md" src="assets/audios/游戏主播女青年.wav" controls></audio>935 </td>936 </tr>937 <tr class="border-bottom-thin">938 <td>939 English-speaking female voice, sounding relatively young,<br>940 with a sweet and pleasant tone. Speaking at a moderate pace<br>941 with a touch of energy, similar to someone narrating a<br>942 beauty/makeup tutorial video. The overall atmosphere is<br>943 relaxed and cheerful.944 </td>945 <td>946 Hi everyone! Today I'll be sharing a soft, romantic<br>947 makeup look that's perfect for dates. Many of you have <br>948 been asking how to apply this eyeshadow naturally - the<br>949 key is using gentle techniques. Let's go through the<br>950 steps together...951 </td>952 <td>953 <audio class="audio-md" src="assets/audios/美妆女博主.wav" controls></audio>954 </td>955 </tr>956 <tr>957 <td>958 English-speaking middle-aged male voice, slightly husky, <br>959 speaking at a moderate-to-slow pace with a deep tone. Like<br>960 someone telling an old story, conveying a nostalgic feeling,<br>961 with a relaxed and composed manner of speaking.962 </td>963 <td>964 That was back in the late 1970s. I remember when our <br>965 village first got electricity - everyone was so excited. <br>966 In theevenings, people would bring their stools and <br>967 gather under the big banyan tree by the village committee <br>968 office to watch movies projected on the wall. Even now, <br>969 thinking back to those moments still fills me with warmth.970 </td>971 <td>972 <audio class="audio-md" src="assets/audios/回忆男中年.wav" controls></audio>973 </td>974 </tr>975 </tbody>976 </table>977 </div>978 </div>979 980 <div class="article-block">981 <h2 id="comparison-of-voice-naturalness">Comparison of voice naturalness982 with the previous generation products</h2>983 <p>The new model demonstrates significant advantages in naturalness compared to the previous version.</p>984 <h3 style="margin-top: 2rem;">Source Audio for Radiant_Girl</h3>985 <audio src="assets/audios/English_Radiant_Girl_Sourse.wav" controls></audio>986 <h3>DEMO</h3>987 <div class="scroll-wrapper">988 <table style="width: 100%;">989 <tbody>990 <tr class="border-bottom-thin">991 <th scope="col">Text</th>992 <th scope="col" style="text-align: center;">MiniMax<br>Speech_02_HD</th>993 <th scope="col" style="text-align: center;">Microsoft<br>Azure TTS</th>994 <th scope="col" style="text-align: center;">AWS<br>Polly</th>995 </tr>996 <tr class="border-bottom-thin">997 <td>998 I sat alone in the empty room, staring at the old photographs,<br>999 wondering how everything could change so quickly,<br>1000 how a lifetime of memories could fade away just like that.1001 </td>1002 <td>1003 <audio class="audio-md" src="assets/audios/Radiant_Girl_1.mp3" controls></audio>1004 </td>1005 <td>1006 <audio class="audio-md" src="assets/audios/Emma_1.mp3" controls></audio>1007 </td>1008 <td>1009 <audio class="audio-md" src="assets/audios/Joanna_1.mp3" controls></audio>1010 </td>1011 </tr>1012 <tr class="border-bottom-thin">1013 <td>1014 The moment I held my acceptance letter, my heart burst with joy - <br>1015 all those sleepless nights finally paid off, and I couldn't stop<br>1016 dancing around the room, calling everyone I knew to share this amazing news!1017 </td>1018 <td>1019 <audio class="audio-md" src="assets/audios/Radiant_Girl_2.mp3" controls></audio>1020 </td>1021 <td>1022 <audio class="audio-md" src="assets/audios/Emma_2.mp3" controls></audio>1023 </td>1024 <td>1025 <audio class="audio-md" src="assets/audios/Joanna_2.mp3" controls></audio>1026 </td>1027 </tr>1028 </tbody>1029 </table>1030 </div>1031 </div>1032 1033 <div class="article-block">1034 <h2 id="citation">Citation</h2>1035 <div>1036 <pre>1037 <code>1038@misc{minimax2025minimaxspeechintrinsiczeroshottexttospeech,1039 title={MiniMax-Speech: Intrinsic Zero-Shot Text-to-Speech with a Learnable Speaker Encoder},1040 author={Bowen Zhang, Congchao Guo, Geng Yang, Hang Yu, Haozhe Zhang, Heidi Lei, Jialong Mai, Junjie Yan, Kaiyue Yang,1041 Mingqi Yang, Peikai Huang, Ruiyang Jin, Sitan Jiang, Weihua Cheng, Yawei Li, Yichen Xiao, Yiying Zhou, Yongmao Zhang,1042 Yuan Lu, Yucen He},1043 year={2025},1044 eprint={2505.07916},1045 archivePrefix={arXiv},1046 primaryClass={eess.AS},1047 url={https://arxiv.org/abs/2505.07916},1048}</code>1049 </pre>1050 </div>1051 </div>1052 </article>1053 </main>1054 1055 <script>1056 MathJax = {1057 tex: {1058 inlineMath: [['$', '$'],],1059 },1060 }1061 1062 const darkModeToggle = document.getElementById('dark-mode-toggle')1063 darkModeToggle.addEventListener('click', () => {1064 document.body.classList.toggle('latex-dark')1065 })1066 </script>1067 <script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>1068</body>1069 1070</html>