--- title: Custom Voices With Qwen3-TTS | daily.dev description: Qwen3-TTS voice embeddings can be cloned, blended and manipulated with vector math to build custom game narrator voices. Qwen3 ships three models: Base for... canonical: https://daily.dev/posts/custom-voices-with-qwen3-tts-vfvsphtod twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Custom Voices With Qwen3-TTS | daily.dev og:description: Qwen3-TTS voice embeddings can be cloned, blended and manipulated with vector math to build custom game narrator voices. Qwen3 ships three models: Base for... og:url: https://daily.dev/posts/custom-voices-with-qwen3-tts-vfvsphtod og:image: https://api.daily.dev/og/posts/vFvSPhTOD.png og:image:alt: Custom Voices With Qwen3-TTS og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Custom Voices With Qwen3-TTS **[Armin Ronacher](https://daily.dev/sources/mitsuhiko)** · 8 min read · 1 upvotes · 0 comments ## Summary Qwen3-TTS voice embeddings can be cloned, blended and manipulated with vector math to build custom game narrator voices. Qwen3 ships three models: Base for cloning from reference audio, CustomVoice with nine built-in speakers plus instructions, and VoiceDesign for inventing voices from text. Base embeddings (2048 numbers) can be weighted-averaged to blend two voices, baked into a new reference clip, or fed into CustomVoice to add emotional instructions like 'speak sad' to a cloned voice. A gender axis built from 14 male and 14 female VoiceDesign voices lets you shift any voice's embedding along that direction. For speaker recognition, a 20-second self-introduction clip got 98% of the other windows right for the same speaker, while a dedicated speaker verification model, ECAPA, found the author in 97% of windows from a conference recording. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How do I clone a voice with Qwen3-TTS and still control emotion with instructions like 'speak sad'? Take the speaker x-vector embedding produced by the Qwen3-TTS Base model's create_voice_clone_prompt function, then pass that embedding into the CustomVoice model's generate call via voice_clone_prompt instead of one of its nine built-in speakers. CustomVoice supports an instruct parameter for tone, so a cloned voice can be told to 'speak in a sad, sorrowful tone,' though instructions work less reliably on cloned voices than on built-in speakers. _Trying to add emotional control to cloned voices? daily.dev surfaces hands-on TTS techniques like this one._ ### Can I blend two voice embeddings together to create a new custom voice in Qwen3-TTS? Yes, speaker embeddings from Qwen3-TTS Base are just 2048-number vectors, so a weighted average of two embeddings produces a voice somewhere between them. A blend function computes (1-t)*a + t*b and renormalizes the result; blending at various ratios retains some features of the original speaker while quickly diverging from sounding exactly like them, and blends work best when the two source recordings have similar spectral/recording characteristics. _Experimenting with voice blending for a project? daily.dev tracks practical write-ups on custom TTS workflows._ ### How accurate is speaker identification using Qwen3-TTS embeddings versus a dedicated speaker verification model like ECAPA? Qwen3-TTS embeddings work poorly for speaker recognition when the reference clip is a noisy conference recording, failing entirely, but using a clean 20-second self-introduction clip as reference correctly identifies the same speaker in 98% of 2-second windows from a podcast episode. A dedicated speaker verification model, ECAPA (speechbrain/spkrec-ecapa-voxceleb), performs better overall, correctly picking out the speaker from the noisy conference clip in 97% of windows. _Choosing between embedding types for speaker diarization? daily.dev helps you compare real benchmark results._ ## Similar posts on daily.dev - [Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning](https://daily.dev/posts/open-tts-leaderboard-scalable-evaluation-for-multilingual-text-to-speech-and-voice-cloning-utmqtcomu) · Hugging Face · 0 upvotes · 0 comments - [Pocket TTS: A high quality TTS that gives your CPU a voice](https://daily.dev/posts/pocket-tts-a-high-quality-tts-that-gives-your-cpu-a-voice-vig6uubge) · Hacker News · 2 upvotes · 0 comments --- Tags: [#python](https://daily.dev/tags/python), [#text-to-speech](https://daily.dev/tags/text-to-speech), [#voice-cloning](https://daily.dev/tags/voice-cloning) [View this post on daily.dev](https://daily.dev/posts/custom-voices-with-qwen3-tts-vfvsphtod) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Custom Voices With Qwen3-TTS","url":"https://daily.dev/posts/custom-voices-with-qwen3-tts-vfvsphtod","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/custom-voices-with-qwen3-tts-vfvsphtod"},"datePublished":"2026-10-29T19:20:41.608Z","dateModified":"2026-10-29T19:21:15.239Z","description":"Qwen3-TTS voice embeddings can be cloned, blended and manipulated with vector math to build custom game narrator voices. Qwen3 ships three models: Base for...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/40a2a0e3c483e6239e620b1bdab5bf69?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/40a2a0e3c483e6239e620b1bdab5bf69?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Armin Ronacher","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Armin Ronacher","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/a7ee3aca0fbd455ea7510f79c0dd61fe","url":"https://daily.dev/sources/mitsuhiko"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/custom-voices-with-qwen3-tts-vfvsphtod","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"python,text-to-speech,voice-cloning","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Armin Ronacher","item":"https://daily.dev/sources/mitsuhiko"},{"@type":"ListItem","position":3,"name":"Custom Voices With Qwen3-TTS"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/custom-voices-with-qwen3-tts-vfvsphtod#faq","mainEntity":[{"@type":"Question","name":"How do I clone a voice with Qwen3-TTS and still control emotion with instructions like 'speak sad'?","acceptedAnswer":{"@type":"Answer","text":"Take the speaker x-vector embedding produced by the Qwen3-TTS Base model's create_voice_clone_prompt function, then pass that embedding into the CustomVoice model's generate call via voice_clone_prompt instead of one of its nine built-in speakers. CustomVoice supports an instruct parameter for tone, so a cloned voice can be told to 'speak in a sad, sorrowful tone,' though instructions work less reliably on cloned voices than on built-in speakers. Trying to add emotional control to cloned voices? daily.dev surfaces hands-on TTS techniques like this one."}},{"@type":"Question","name":"Can I blend two voice embeddings together to create a new custom voice in Qwen3-TTS?","acceptedAnswer":{"@type":"Answer","text":"Yes, speaker embeddings from Qwen3-TTS Base are just 2048-number vectors, so a weighted average of two embeddings produces a voice somewhere between them. A blend function computes (1-t)*a + t*b and renormalizes the result; blending at various ratios retains some features of the original speaker while quickly diverging from sounding exactly like them, and blends work best when the two source recordings have similar spectral/recording characteristics. Experimenting with voice blending for a project? daily.dev tracks practical write-ups on custom TTS workflows."}},{"@type":"Question","name":"How accurate is speaker identification using Qwen3-TTS embeddings versus a dedicated speaker verification model like ECAPA?","acceptedAnswer":{"@type":"Answer","text":"Qwen3-TTS embeddings work poorly for speaker recognition when the reference clip is a noisy conference recording, failing entirely, but using a clean 20-second self-introduction clip as reference correctly identifies the same speaker in 98% of 2-second windows from a podcast episode. A dedicated speaker verification model, ECAPA (speechbrain/spkrec-ecapa-voxceleb), performs better overall, correctly picking out the speaker from the noisy conference clip in 97% of windows. Choosing between embedding types for speaker diarization? daily.dev helps you compare real benchmark results."}}]} ```