Add Higgs Audio v3 TTS engine on Apple Silicon (MLX)

.github/workflows/release.yml

@@ -101,7 +101,7 @@ jobs:
           # rest (sounddevice, protobuf, sentencepiece, pyyaml, jinja2) are
           # pulled in by other engines.
           pip install --no-deps mlx-lm==0.31.1
-          pip install --no-deps mlx-audio==0.4.1
+          pip install --no-deps mlx-audio==0.5.9
 
       - name: Build Python server (Linux/macOS)
         if: matrix.platform != 'windows-latest'

CHANGELOG.md

@@ -31,6 +31,7 @@
 ### Engines
 
 - **OmniVoice TTS engine (non-commercial).** Adds k2-fsa/OmniVoice as an opt-in eighth engine: zero-shot cloning and attribute-based voice design (gender, age, pitch, accent) across the 23 languages Voicebox already lists, 24 kHz output. Its `HiggsAudioV2` codec only exists in transformers 5.x, so the modelling code is vendored under `backend/vendor/` and grafted onto the pinned 4.57 at load time; `omnivoice` installs `--no-deps` like Chatterbox and TADA. The weights are CC-BY-NC, so the engine is labelled non-commercial in the engine picker and model manager and is never a default. — @pippo73 ([#1062](https://github.com/jamiepine/voicebox/pull/1062); fixes [#380](https://github.com/jamiepine/voicebox/issues/380), [#517](https://github.com/jamiepine/voicebox/issues/517), [#791](https://github.com/jamiepine/voicebox/issues/791))
+- **Higgs Audio v3 TTS engine on Apple Silicon (non-commercial).** Adds Boson AI's Higgs Audio v3 4B (`bosonai/higgs-tts-3-4b`) as the `higgs` engine, MLX only: zero-shot cloning across the 23 languages Voicebox lists, inline control tags such as `<|emotion:amusement|>`, 24 kHz output, about 0.9x realtime on an M4 Max. It runs on mlx-audio's port, so mlx-audio moves from 0.4.1 to 0.5.9 (still `--no-deps`; mlx now needs 0.31.1+). The bump changed Whisper's temperature fallback into multilingual gibberish on short clips, so MLX transcription now decodes greedily. The community 4-bit MLX artifact (`Reza2kn/Higgs-Audio-v3-TTS-4bit-MLX`) holds only the quantized transformer layers, without the audio codec, so it cannot be loaded; the engine downloads the 9.3 GB bf16 checkpoint instead. The weights are under Boson's research and non-commercial licence, so the engine is labelled non-commercial and is never a default.
 
 ## [0.6.0] - 2026-10-04
 

app/src/components/CapturesTab/CapturesTab.tsx

@@ -269,7 +269,7 @@ export function CapturesTab() {
       // override fall through to whatever the backend picks.
       const engine = voice.default_engine as
         | 'qwen' | 'qwen_custom_voice' | 'qwen_voice_design' | 'luxtts' | 'chatterbox'
-        | 'chatterbox_turbo' | 'tada' | 'kokoro' | 'omnivoice'
+        | 'chatterbox_turbo' | 'tada' | 'kokoro' | 'omnivoice' | 'higgs'
         | undefined;
       return apiClient.generateSpeech({
         profile_id: voice.id,

app/src/components/Generation/EngineModelSelector.tsx

@@ -29,6 +29,7 @@ const ENGINE_OPTIONS = [
   { value: 'tada:3B', label: 'TADA 3B Multilingual', engine: 'tada' },
   { value: 'kokoro', label: 'Kokoro 82M', engine: 'kokoro' },
   { value: 'omnivoice', label: 'OmniVoice (non-commercial)', engine: 'omnivoice' },
+  { value: 'higgs', label: 'Higgs Audio v3 (non-commercial)', engine: 'higgs' },
 ] as const;
 
 const ENGINE_DESCRIPTIONS: Record<string, string> = {
@@ -41,6 +42,7 @@ const ENGINE_DESCRIPTIONS: Record<string, string> = {
   tada: 'HumeAI, 700s+ coherent audio',
   kokoro: '82M params, CPU realtime, 8 langs',
   omnivoice: '23 languages, voice design. CC-BY-NC weights: non-commercial use only',
+  higgs: 'Apple Silicon only, 23 languages, <|emotion:...|> tags. Non-commercial licence',
 };
 
 /** Engines that only support English and should force language to 'en' on select. */
@@ -54,6 +56,7 @@ const CLONING_ENGINES = new Set([
   'chatterbox_turbo',
   'tada',
   'omnivoice',
+  'higgs',
 ]);
 
 /** Engines that synthesise a voice from a written description. */

app/src/components/Generation/FloatingGenerateBox.tsx

@@ -154,7 +154,8 @@ export function FloatingGenerateBox({
     | 'kokoro'
     | 'qwen_custom_voice'
     | 'qwen_voice_design'
-    | 'omnivoice';
+    | 'omnivoice'
+    | 'higgs';
   useEffect(() => {
     if (selectedProfile?.language) {
       form.setValue('language', selectedProfile.language as LanguageCode);

app/src/components/ServerSettings/ModelManagement.tsx

@@ -71,6 +71,8 @@ const MODEL_DESCRIPTIONS: Record<string, string> = {
     'Kokoro 82M by hexgrad. Tiny 82M-parameter TTS that runs at CPU realtime. Supports 8 languages with pre-built voice styles. Apache 2.0 licensed.',
   omnivoice:
     'OmniVoice by k2-fsa. Zero-shot voice cloning via a diffusion language model, plus attribute-based voice design (gender, age, pitch, accent). The model covers 600+ languages; Voicebox exposes the 23 in its own language list. 24kHz output. NON-COMMERCIAL: the weights are CC-BY-NC (code is Apache 2.0; the bundled Higgs Audio V2 codec is under the Boson community licence), so do not use output commercially.',
+  'higgs-audio-v3':
+    'Higgs Audio v3 4B by Boson AI, on Apple Silicon (MLX). Qwen3-based conversational TTS with zero-shot voice cloning, 100+ languages (Voicebox exposes 23) and inline control tags such as <|emotion:amusement|> or <|sfx:laughter|>. 24kHz output, ~9.3 GB download. NON-COMMERCIAL: Boson Higgs Audio v3 research and non-commercial licence.',
   'qwen-custom-voice-1.7B':
     'Qwen3-TTS CustomVoice 1.7B by Alibaba. 9 premium preset voices with instruct-based style control for tone, emotion, and prosody. Supports 10 languages.',
   'qwen-custom-voice-0.6B':
@@ -420,7 +422,8 @@ export function ModelManagement() {
         m.model_name.startsWith('chatterbox') ||
         m.model_name.startsWith('tada') ||
         m.model_name.startsWith('kokoro') ||
-        m.model_name.startsWith('omnivoice'),
+        m.model_name.startsWith('omnivoice') ||
+        m.model_name.startsWith('higgs'),
     ) ?? [];
   const whisperModels = modelStatus?.models.filter((m) => m.model_name.startsWith('whisper')) ?? [];
   const llmModels = modelStatus?.models.filter((m) => m.model_name.startsWith('qwen3-')) ?? [];

app/src/components/VoiceProfiles/ProfileForm.tsx

@@ -76,6 +76,7 @@ const DEFAULT_ENGINE_OPTIONS = [
   { value: 'tada', label: 'TADA' },
   { value: 'kokoro', label: 'Kokoro 82M' },
   { value: 'omnivoice', label: 'OmniVoice (non-commercial)' },
+  { value: 'higgs', label: 'Higgs Audio v3 (non-commercial)' },
 ] as const;
 
 function makeProfileSchema(t: (key: string) => string) {

app/src/lib/api/types.ts

@@ -80,7 +80,8 @@ export interface GenerationRequest {
     | 'chatterbox_turbo'
     | 'tada'
     | 'kokoro'
-    | 'omnivoice';
+    | 'omnivoice'
+    | 'higgs';
   instruct?: string;
   /** When true and the profile has a personality prompt, input text is rewritten in-character before TTS. */
   personality?: boolean;

app/src/lib/constants/languages.ts

@@ -98,6 +98,32 @@ export const ENGINE_LANGUAGES: Record<string, readonly LanguageCode[]> = {
     'tr',
     'zh',
   ],
+  // Higgs Audio v3 covers 100+ languages; same ALL_LANGUAGES subset as OmniVoice.
+  higgs: [
+    'ar',
+    'da',
+    'de',
+    'el',
+    'en',
+    'es',
+    'fi',
+    'fr',
+    'he',
+    'hi',
+    'it',
+    'ja',
+    'ko',
+    'ms',
+    'nl',
+    'no',
+    'pl',
+    'pt',
+    'ru',
+    'sv',
+    'sw',
+    'tr',
+    'zh',
+  ],
 } as const;
 
 /** Helper: get language options for a given engine. */

app/src/lib/hooks/useGenerationForm.ts

@@ -30,6 +30,7 @@ const generationSchema = z.object({
       'tada',
       'kokoro',
       'omnivoice',
+      'higgs',
     ])
     .optional(),
   personality: z.boolean().optional(),
@@ -107,9 +108,11 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
                     ? 'qwen-voice-design-1.7B'
                     : engine === 'omnivoice'
                       ? 'omnivoice'
-                      : engine === 'qwen_custom_voice'
-                        ? `qwen-custom-voice-${data.modelSize}`
-                        : `qwen-tts-${data.modelSize}`;
+                      : engine === 'higgs'
+                        ? 'higgs-audio-v3'
+                        : engine === 'qwen_custom_voice'
+                          ? `qwen-custom-voice-${data.modelSize}`
+                          : `qwen-tts-${data.modelSize}`;
       const displayName =
         engine === 'luxtts'
           ? 'LuxTTS'
@@ -127,13 +130,15 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
                     ? 'Qwen VoiceDesign 1.7B'
                     : engine === 'omnivoice'
                       ? 'OmniVoice'
-                      : engine === 'qwen_custom_voice'
-                        ? data.modelSize === '1.7B'
-                          ? 'Qwen CustomVoice 1.7B'
-                          : 'Qwen CustomVoice 0.6B'
-                        : data.modelSize === '1.7B'
-                          ? 'Qwen TTS 1.7B'
-                          : 'Qwen TTS 0.6B';
+                      : engine === 'higgs'
+                        ? 'Higgs Audio v3'
+                        : engine === 'qwen_custom_voice'
+                          ? data.modelSize === '1.7B'
+                            ? 'Qwen CustomVoice 1.7B'
+                            : 'Qwen CustomVoice 0.6B'
+                          : data.modelSize === '1.7B'
+                            ? 'Qwen TTS 1.7B'
+                            : 'Qwen TTS 0.6B';
 
       // Check if model needs downloading
       try {

app/src/lib/utils/format.ts

@@ -64,6 +64,7 @@ const ENGINE_DISPLAY_NAMES: Record<string, string> = {
   chatterbox_turbo: 'Chatterbox Turbo',
   tada: 'TADA',
   kokoro: 'Kokoro 82M',
+  higgs: 'Higgs Audio v3',
 };
 
 export function formatEngineName(engine?: string, modelSize?: string): string {

backend/backends/__init__.py

@@ -239,6 +239,7 @@ def is_loaded(self) -> bool:
     "tada": "TADA",
     "kokoro": "Kokoro",
     "omnivoice": "OmniVoice",
+    "higgs": "Higgs Audio v3",
 }
 
 LLM_ENGINES = {
@@ -347,7 +348,49 @@ def _get_non_qwen_tts_configs() -> list[ModelConfig]:
         chatterbox_size_mb = 3200
         chatterbox_aux_repos = ()
 
-    return [
+    # Higgs Audio v3 only has an MLX implementation (mlx-audio), so it is offered on Apple Silicon only.
+    higgs_configs = (
+        [
+            ModelConfig(
+                model_name="higgs-audio-v3",
+                # Non-commercial: Boson Higgs Audio v3 research and non-commercial licence.
+                display_name="Higgs Audio v3 4B (Multilingual, non-commercial)",
+                engine="higgs",
+                hf_repo_id="bosonai/higgs-tts-3-4b",
+                size_mb=9300,
+                # Same 23 languages Voicebox already exposes; upstream covers 100+.
+                languages=[
+                    "zh",
+                    "en",
+                    "ja",
+                    "ko",
+                    "de",
+                    "fr",
+                    "ru",
+                    "pt",
+                    "es",
+                    "it",
+                    "he",
+                    "ar",
+                    "da",
+                    "el",
+                    "fi",
+                    "hi",
+                    "ms",
+                    "nl",
+                    "no",
+                    "pl",
+                    "sv",
+                    "sw",
+                    "tr",
+                ],
+            )
+        ]
+        if on_mlx
+        else []
+    )
+
+    return higgs_configs + [
         ModelConfig(
             model_name="luxtts",
             display_name="LuxTTS (Fast, CPU-friendly)",
@@ -842,6 +885,12 @@ def get_tts_backend_for_engine(engine: str) -> TTSBackend:
             from .omnivoice_backend import OmniVoiceBackend
 
             backend = OmniVoiceBackend()
+        elif engine == "higgs":
+            if get_backend_type() != "mlx":
+                raise ValueError("Higgs Audio v3 requires Apple Silicon (MLX backend)")
+            from .higgs_mlx_backend import HiggsMLXTTSBackend
+
+            backend = HiggsMLXTTSBackend()
         elif engine == "qwen_custom_voice":
             from .qwen_custom_voice_backend import QwenCustomVoiceBackend