OFFLINE AI NPCQwen speech: language, voice and style
In 1.1, choose a ready CustomVoice speaker. Creating your own voice from a description (VoiceDesign) is not part of this version.
For the current Unity Qwen3-TTS component, start with a CustomVoice talker and leave Style empty. For the overall model workflow, see Current model usage. These settings have separate jobs:
| Setting | What it controls | Example |
|---|---|---|
| Text | Words to say | Hello. I am ready. |
| Language | Language of those words | Unity code en; runtime label english |
| Voice | A named preset timbre | ryan |
| Style | Delivery instruction, where the model supports it | Speak calmly, at a measured pace. |
Describe delivery in Style; set the language separately. A style such as “Speak Russian” does not replace the language setting or translate the text.
Which model accepts which input?
| Qwen3-TTS variant | Voice input | Instructions |
|---|---|---|
| 0.6B CustomVoice | Named preset speaker | Instruction control is unsupported; leave Style empty |
| 1.7B CustomVoice | Named preset speaker | Optional delivery instruction |
| 1.7B VoiceDesign | Description of a voice | A separate VoiceDesign operation |
| 0.6B / 1.7B Base | Reference audio or a prepared clone prompt | A separate cloning operation |
The bundled native transport explicitly accepts CustomVoice; Base and VoiceDesign are not configured. The optional HTTP request shape also carries a named voice, so it does not provide a VoiceDesign or reference-audio authoring flow. See the official model overview, official Python API.
In Unity
using OfflineAINPC.Speech;
// CustomVoice, with no style dependency:
var line = new SpeechRequest("Hello. I am ready.",
languageCode: "en", voiceId: "ryan");
// For a separately verified 1.7B CustomVoice setup:
var styledLine = new SpeechRequest("Hello. I am ready.",
languageCode: "en", voiceId: "ryan",
style: "Speak calmly, at a measured pace.");An explicit SpeechRequest.LanguageCode overrides the engine language. Native codes en, ru, es, zh, ja, ko, de, fr, pt, it accept region/script suffixes and _; blank or unknown codes select auto. This is language selection, not translation. The native DLL passes text, language, speaker and optional style directly to the pinned runtime. HTTP field omission/default behavior belongs only to an explicitly configured external server. Leave both engine Style and request Style empty for 0.6B CustomVoice. A 1.7B instruction setup requires its own model and runtime check.
Model Manager downloads the speech model files. The package includes a Windows x64 CPU Qwen native DLL; leave External server URL empty to use it. The ordinary Qwen route needs no manually started HTTP server, executable or compiler setup. Keep an existing working Piper route. See the native speech setup and Licensing.