Merge pull request #222 from soniqo/add/madlad-translation

Add MADLADTranslation: 400+ language translation (T5 v1.1, MLX)

Author: ivan-digital

Package.swift

@@ -76,6 +76,10 @@ let package = Package(
             name: "Qwen3Chat",
             targets: ["Qwen3Chat"]
         ),
+        .library(
+            name: "MADLADTranslation",
+            targets: ["MADLADTranslation"]
+        ),
         .library(
             name: "SpeechUI",
             targets: ["SpeechUI"]
@@ -260,6 +264,17 @@ let package = Package(
                 .product(name: "MLXFast", package: "mlx-swift"),
             ]
         ),
+        .target(
+            name: "MADLADTranslation",
+            dependencies: [
+                "AudioCommon",
+                "MLXCommon",
+                .product(name: "MLX", package: "mlx-swift"),
+                .product(name: "MLXNN", package: "mlx-swift"),
+                .product(name: "MLXFast", package: "mlx-swift"),
+                .product(name: "Transformers", package: "swift-transformers"),
+            ]
+        ),
         .target(
             name: "SpeechUI",
             dependencies: []
@@ -285,6 +300,7 @@ let package = Package(
                 "OmnilingualASR",
                 "KokoroTTS",
                 "VibeVoiceTTS",
+                "MADLADTranslation",
                 "SpeechWakeWord",
                 "AudioCommon",
                 .product(name: "MLX", package: "mlx-swift"),
@@ -425,6 +441,13 @@ let package = Package(
                 "AudioCommon",
             ]
         ),
+        .testTarget(
+            name: "MADLADTranslationTests",
+            dependencies: [
+                "MADLADTranslation",
+                "AudioCommon",
+            ]
+        ),
         .testTarget(
             name: "AudioCLITests",
             dependencies: [

README.md

@@ -6,7 +6,7 @@ AI speech models for Apple Silicon, powered by MLX Swift and CoreML.
 
 On-device speech recognition, synthesis, and understanding for Mac and iOS. Runs locally on Apple Silicon — no cloud, no API keys, no data leaves your device.
 
-**[📚 Full Documentation →](https://soniqo.audio)** · **[🤗 HuggingFace Models](https://huggingface.co/aufklarer)** · **[📝 Blog](https://blog.ivan.digital)**
+**[📚 Full Documentation →](https://soniqo.audio)** · **[🤗 HuggingFace Models](https://huggingface.co/aufklarer)** · **[📝 Blog](https://blog.ivan.digital)** · **[💬 Discord](https://discord.gg/TnCryqEMgu)**
 
 - **[Qwen3-ASR](https://soniqo.audio/guides/transcribe)** — Speech-to-text (automatic speech recognition, 52 languages, MLX + CoreML)
 - **[Parakeet TDT](https://soniqo.audio/guides/parakeet)** — Speech-to-text via CoreML (Neural Engine, NVIDIA FastConformer + TDT decoder, 25 languages)
@@ -19,6 +19,7 @@ On-device speech recognition, synthesis, and understanding for Mac and iOS. Runs
 - **[Kokoro TTS](https://soniqo.audio/guides/kokoro)** — On-device TTS (82M, CoreML/Neural Engine, 54 voices, iOS-ready, 10 languages)
 - **[VibeVoice TTS](https://soniqo.audio/guides/vibevoice)** — Long-form / multi-speaker TTS (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, up to 90-min podcast/audiobook synthesis, EN/ZH)
 - **[Qwen3.5-Chat](https://soniqo.audio/guides/chat)** — On-device LLM chat (0.8B, MLX INT4 + CoreML INT8, DeltaNet hybrid, streaming tokens)
+- **[MADLAD-400](https://soniqo.audio/guides/translate)** — Many-to-many translation across 400+ languages (3B, MLX INT4 + INT8, T5 v1.1, Apache 2.0)
 - **[PersonaPlex](https://soniqo.audio/guides/respond)** — Full-duplex speech-to-speech (7B, audio in → audio out, 18 voice presets)
 - **[DeepFilterNet3](https://soniqo.audio/guides/denoise)** — Real-time noise suppression (2.1M params, 48 kHz)
 - **[Source Separation](https://soniqo.audio/guides/separate)** — Music source separation via Open-Unmix (UMX-HQ / UMX-L, 4 stems: vocals/drums/bass/other, 44.1 kHz stereo)
@@ -115,6 +116,7 @@ Compact view below. **[Full model catalogue with sizes, quantisations, download
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/guides/vibevoice) | Text → Speech (long-form, multi-speaker) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/guides/vibevoice) | Text → Speech (up to 90-min podcast) | MLX | 1.5B | EN/ZH |
 | [Qwen3.5-Chat](https://soniqo.audio/guides/chat) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
+| [MADLAD-400](https://soniqo.audio/guides/translate) | Text → Text (Translation) | MLX | 3B | **400+** |
 | [PersonaPlex](https://soniqo.audio/guides/respond) | Speech → Speech | MLX | 7B | EN |
 | [Silero VAD](https://soniqo.audio/guides/vad) | Voice Activity Detection | MLX, CoreML | 309K | Agnostic |
 | [Pyannote](https://soniqo.audio/guides/diarize) | VAD + Diarization | MLX | 1.5M | Agnostic |
@@ -139,6 +141,7 @@ Then:
 ```bash
 audio transcribe recording.wav
 audio speak "Hello world"
+audio translate "Hello, how are you?" --to es
 audio respond --input question.wav --transcript
 audio-server --port 8080            # local HTTP / WebSocket server (OpenAI-compatible /v1/realtime)
 ```
@@ -166,6 +169,7 @@ import CosyVoiceTTS         // Text-to-speech with voice cloning
 import KokoroTTS            // Text-to-speech (iOS-ready)
 import VibeVoiceTTS         // Long-form / multi-speaker TTS (EN/ZH)
 import Qwen3Chat            // On-device LLM chat
+import MADLADTranslation    // Many-to-many translation across 400+ languages
 import PersonaPlex          // Full-duplex speech-to-speech
 import SpeechVAD            // VAD + speaker diarization + embeddings
 import SpeechEnhancement    // Noise suppression
@@ -267,6 +271,16 @@ chat.chat(messages: [(.user, "Explain MLX in one sentence")]) { token, isFinal i
 }
 ```
 
+### Translation — [full guide →](https://soniqo.audio/guides/translate)
+
+```swift
+import MADLADTranslation
+
+let translator = try await MADLADTranslator.fromPretrained()
+let es = try translator.translate("Hello, how are you?", to: "es")
+// → "Hola, ¿cómo estás?"
+```
+
 ### Voice Activity Detection — [full guide →](https://soniqo.audio/guides/vad)
 
 ```swift

README_de.md

@@ -6,7 +6,7 @@ KI-Sprachmodelle für Apple Silicon, basierend auf MLX Swift und CoreML.
 
 Spracherkennung, -synthese und -verständnis auf dem Gerät für Mac und iOS. Läuft vollständig lokal auf Apple Silicon — keine Cloud, keine API-Schlüssel, keine Daten verlassen das Gerät.
 
-**[📚 Vollständige Dokumentation →](https://soniqo.audio/de)** · **[🤗 HuggingFace-Modelle](https://huggingface.co/aufklarer)** · **[📝 Blog](https://blog.ivan.digital)**
+**[📚 Vollständige Dokumentation →](https://soniqo.audio/de)** · **[🤗 HuggingFace-Modelle](https://huggingface.co/aufklarer)** · **[📝 Blog](https://blog.ivan.digital)** · **[💬 Discord](https://discord.gg/TnCryqEMgu)**
 
 - **[Qwen3-ASR](https://soniqo.audio/de/guides/transcribe)** — Sprache-zu-Text (automatische Spracherkennung, 52 Sprachen, MLX + CoreML)
 - **[Parakeet TDT](https://soniqo.audio/de/guides/parakeet)** — Sprache-zu-Text über CoreML (Neural Engine, NVIDIA FastConformer + TDT-Decoder, 25 Sprachen)
@@ -19,6 +19,7 @@ Spracherkennung, -synthese und -verständnis auf dem Gerät für Mac und iOS. L
 - **[Kokoro TTS](https://soniqo.audio/de/guides/kokoro)** — TTS auf dem Gerät (82M, CoreML/Neural Engine, 54 Stimmen, iOS-tauglich, 10 Sprachen)
 - **[VibeVoice TTS](https://soniqo.audio/de/guides/vibevoice)** — Langform-/Multi-Speaker-TTS (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, bis zu 90 Min. Podcast-/Hörbuch-Synthese, EN/ZH)
 - **[Qwen3.5-Chat](https://soniqo.audio/de/guides/chat)** — LLM-Chat auf dem Gerät (0.8B, MLX INT4 + CoreML INT8, DeltaNet-Hybrid, Token-Streaming)
+- **[MADLAD-400](https://soniqo.audio/de/guides/translate)** — Mehrsprachige Übersetzung über 400+ Sprachen (3B, MLX INT4 + INT8, T5 v1.1, Apache 2.0)
 - **[PersonaPlex](https://soniqo.audio/de/guides/respond)** — Vollduplex-Sprache-zu-Sprache (7B, Audio rein → Audio raus, 18 Stimmvoreinstellungen)
 - **[DeepFilterNet3](https://soniqo.audio/de/guides/denoise)** — Echtzeit-Rauschunterdrückung (2,1M Parameter, 48 kHz)
 - **[Quelltrennung](https://soniqo.audio/de/guides/separate)** — Musikquelltrennung mit Open-Unmix (UMX-HQ / UMX-L, 4 Stems: Gesang/Drums/Bass/Rest, 44,1 kHz Stereo)
@@ -115,6 +116,7 @@ Kompakte Übersicht unten. **[Vollständiger Modellkatalog mit Größen, Quantis
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/de/guides/vibevoice) | Text → Sprache (Langform, Multi-Speaker) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/de/guides/vibevoice) | Text → Sprache (bis zu 90 Min. Podcast) | MLX | 1.5B | EN/ZH |
 | [Qwen3.5-Chat](https://soniqo.audio/de/guides/chat) | Text → Text (LLM) | MLX, CoreML | 0.8B | Multi |
+| [MADLAD-400](https://soniqo.audio/de/guides/translate) | Text → Text (Übersetzung) | MLX | 3B | **400+** |
 | [PersonaPlex](https://soniqo.audio/de/guides/respond) | Sprache → Sprache | MLX | 7B | EN |
 | [Silero VAD](https://soniqo.audio/de/guides/vad) | Sprachaktivitätserkennung | MLX, CoreML | 309K | Sprachunabhängig |
 | [Pyannote](https://soniqo.audio/de/guides/diarize) | VAD + Diarisierung | MLX | 1.5M | Sprachunabhängig |
@@ -139,6 +141,7 @@ Dann:
 ```bash
 audio transcribe recording.wav
 audio speak "Hello world"
+audio translate "Hello, how are you?" --to es
 audio respond --input question.wav --transcript
 audio-server --port 8080            # lokaler HTTP/WebSocket-Server (OpenAI-kompatibles /v1/realtime)
 ```
@@ -166,6 +169,7 @@ import CosyVoiceTTS         // Sprachsynthese mit Stimmklonen
 import KokoroTTS            // Sprachsynthese (iOS-tauglich)
 import VibeVoiceTTS         // Langform-/Multi-Speaker-TTS (EN/ZH)
 import Qwen3Chat            // LLM-Chat auf dem Gerät
+import MADLADTranslation    // Mehrsprachige Übersetzung über 400+ Sprachen
 import PersonaPlex          // Vollduplex-Sprache-zu-Sprache
 import SpeechVAD            // VAD + Sprecherdiarisierung + Einbettungen
 import SpeechEnhancement    // Rauschunterdrückung
@@ -267,6 +271,16 @@ chat.chat(messages: [(.user, "Explain MLX in one sentence")]) { token, isFinal i
 }
 ```
 
+### Übersetzung — [vollständige Anleitung →](https://soniqo.audio/de/guides/translate)
+
+```swift
+import MADLADTranslation
+
+let translator = try await MADLADTranslator.fromPretrained()
+let es = try translator.translate("Hello, how are you?", to: "es")
+// → "Hola, ¿cómo estás?"
+```
+
 ### Sprachaktivitätserkennung — [vollständige Anleitung →](https://soniqo.audio/de/guides/vad)
 
 ```swift

README_es.md

@@ -6,7 +6,7 @@ Modelos de IA de voz para Apple Silicon, impulsados por MLX Swift y CoreML.
 
 Reconocimiento, síntesis y comprensión de voz en el dispositivo para Mac e iOS. Se ejecuta localmente en Apple Silicon — sin nube, sin claves de API, ningún dato sale del dispositivo.
 
-**[📚 Documentación completa →](https://soniqo.audio/es)** · **[🤗 Modelos en HuggingFace](https://huggingface.co/aufklarer)** · **[📝 Blog](https://blog.ivan.digital)**
+**[📚 Documentación completa →](https://soniqo.audio/es)** · **[🤗 Modelos en HuggingFace](https://huggingface.co/aufklarer)** · **[📝 Blog](https://blog.ivan.digital)** · **[💬 Discord](https://discord.gg/TnCryqEMgu)**
 
 - **[Qwen3-ASR](https://soniqo.audio/es/guides/transcribe)** — Voz a texto (reconocimiento automático del habla, 52 idiomas, MLX + CoreML)
 - **[Parakeet TDT](https://soniqo.audio/es/guides/parakeet)** — Voz a texto vía CoreML (Neural Engine, NVIDIA FastConformer + decodificador TDT, 25 idiomas)
@@ -19,6 +19,7 @@ Reconocimiento, síntesis y comprensión de voz en el dispositivo para Mac e iOS
 - **[Kokoro TTS](https://soniqo.audio/es/guides/kokoro)** — TTS en el dispositivo (82M, CoreML/Neural Engine, 54 voces, listo para iOS, 10 idiomas)
 - **[VibeVoice TTS](https://soniqo.audio/es/guides/vibevoice)** — TTS de formato largo / múltiples hablantes (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, síntesis de podcast/audiolibro de hasta 90 min, EN/ZH)
 - **[Qwen3.5-Chat](https://soniqo.audio/es/guides/chat)** — Chat LLM en el dispositivo (0.8B, MLX INT4 + CoreML INT8, DeltaNet híbrido, tokens en streaming)
+- **[MADLAD-400](https://soniqo.audio/es/guides/translate)** — Traducción multidireccional entre 400+ idiomas (3B, MLX INT4 + INT8, T5 v1.1, Apache 2.0)
 - **[PersonaPlex](https://soniqo.audio/es/guides/respond)** — Voz a voz full-duplex (7B, audio de entrada → audio de salida, 18 presets de voz)
 - **[DeepFilterNet3](https://soniqo.audio/es/guides/denoise)** — Supresión de ruido en tiempo real (2.1M parámetros, 48 kHz)
 - **[Separación de fuentes](https://soniqo.audio/es/guides/separate)** — Separación de fuentes musicales con Open-Unmix (UMX-HQ / UMX-L, 4 stems: voces/batería/bajo/otros, 44,1 kHz estéreo)
@@ -115,6 +116,7 @@ Vista compacta a continuación. **[Catálogo completo de modelos con tamaños, c
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/es/guides/vibevoice) | Texto → Voz (formato largo, múltiples hablantes) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/es/guides/vibevoice) | Texto → Voz (podcast de hasta 90 min) | MLX | 1.5B | EN/ZH |
 | [Qwen3.5-Chat](https://soniqo.audio/es/guides/chat) | Texto → Texto (LLM) | MLX, CoreML | 0.8B | Multi |
+| [MADLAD-400](https://soniqo.audio/es/guides/translate) | Texto → Texto (Traducción) | MLX | 3B | **400+** |
 | [PersonaPlex](https://soniqo.audio/es/guides/respond) | Voz → Voz | MLX | 7B | EN |
 | [Silero VAD](https://soniqo.audio/es/guides/vad) | Detección de actividad vocal | MLX, CoreML | 309K | Agnóstico |
 | [Pyannote](https://soniqo.audio/es/guides/diarize) | VAD + Diarización | MLX | 1.5M | Agnóstico |
@@ -139,6 +141,7 @@ Luego:
 ```bash
 audio transcribe recording.wav
 audio speak "Hello world"
+audio translate "Hello, how are you?" --to es
 audio respond --input question.wav --transcript
 audio-server --port 8080            # servidor HTTP / WebSocket local (OpenAI-compatible /v1/realtime)
 ```
@@ -166,6 +169,7 @@ import CosyVoiceTTS         // Síntesis de voz con clonación
 import KokoroTTS            // Síntesis de voz (listo para iOS)
 import VibeVoiceTTS         // TTS de formato largo / múltiples hablantes (EN/ZH)
 import Qwen3Chat            // Chat LLM en el dispositivo
+import MADLADTranslation    // Traducción multidireccional entre 400+ idiomas
 import PersonaPlex          // Voz a voz full-duplex
 import SpeechVAD            // VAD + diarización + embeddings
 import SpeechEnhancement    // Supresión de ruido
@@ -267,6 +271,16 @@ chat.chat(messages: [(.user, "Explain MLX in one sentence")]) { token, isFinal i
 }
 ```
 
+### Traducción — [guía completa →](https://soniqo.audio/es/guides/translate)
+
+```swift
+import MADLADTranslation
+
+let translator = try await MADLADTranslator.fromPretrained()
+let es = try translator.translate("Hello, how are you?", to: "es")
+// → "Hola, ¿cómo estás?"
+```
+
 ### Detección de actividad vocal — [guía completa →](https://soniqo.audio/es/guides/vad)
 
 ```swift

README_fr.md

@@ -6,7 +6,7 @@ Modeles IA de parole pour Apple Silicon, propulses par MLX Swift et CoreML.
 
 Reconnaissance, synthese et comprehension vocale embarquees pour Mac et iOS. S'execute entierement en local sur Apple Silicon -- sans cloud, sans cle API, aucune donnee ne quitte l'appareil.
 
-**[📚 Documentation complete →](https://soniqo.audio/fr)** · **[🤗 Modeles HuggingFace](https://huggingface.co/aufklarer)** · **[📝 Blog](https://blog.ivan.digital)**
+**[📚 Documentation complete →](https://soniqo.audio/fr)** · **[🤗 Modeles HuggingFace](https://huggingface.co/aufklarer)** · **[📝 Blog](https://blog.ivan.digital)** · **[💬 Discord](https://discord.gg/TnCryqEMgu)**
 
 - **[Qwen3-ASR](https://soniqo.audio/fr/guides/transcribe)** -- Reconnaissance vocale (reconnaissance automatique de la parole, 52 langues, MLX + CoreML)
 - **[Parakeet TDT](https://soniqo.audio/fr/guides/parakeet)** -- Reconnaissance vocale via CoreML (Neural Engine, NVIDIA FastConformer + decodeur TDT, 25 langues)
@@ -19,6 +19,7 @@ Reconnaissance, synthese et comprehension vocale embarquees pour Mac et iOS. S'e
 - **[Kokoro TTS](https://soniqo.audio/fr/guides/kokoro)** -- TTS embarque (82M, CoreML/Neural Engine, 54 voix, compatible iOS, 10 langues)
 - **[VibeVoice TTS](https://soniqo.audio/fr/guides/vibevoice)** -- TTS long format / multi-locuteurs (Microsoft VibeVoice Realtime-0.5B + 1.5B, MLX, synthese de podcast/livre audio jusqu'a 90 min, EN/ZH)
 - **[Qwen3.5-Chat](https://soniqo.audio/fr/guides/chat)** -- Chat LLM embarque (0.8B, MLX INT4 + CoreML INT8, DeltaNet hybride, tokens en streaming)
+- **[MADLAD-400](https://soniqo.audio/fr/guides/translate)** — Traduction multidirectionnelle entre 400+ langues (3B, MLX INT4 + INT8, T5 v1.1, Apache 2.0)
 - **[PersonaPlex](https://soniqo.audio/fr/guides/respond)** -- Parole-a-parole en full-duplex (7B, audio entrant → audio sortant, 18 preselections de voix)
 - **[DeepFilterNet3](https://soniqo.audio/fr/guides/denoise)** -- Suppression de bruit en temps reel (2,1M parametres, 48 kHz)
 - **[Séparation de sources](https://soniqo.audio/fr/guides/separate)** — Séparation de sources musicales avec Open-Unmix (UMX-HQ / UMX-L, 4 stems : voix/batterie/basse/autres, 44,1 kHz stéréo)
@@ -115,6 +116,7 @@ Vue compacte ci-dessous. **[Catalogue complet des modeles avec tailles, quantifi
 | [VibeVoice Realtime-0.5B](https://soniqo.audio/fr/guides/vibevoice) | Texte → Parole (long format, multi-locuteurs) | MLX | 0.5B | EN/ZH |
 | [VibeVoice 1.5B](https://soniqo.audio/fr/guides/vibevoice) | Texte → Parole (podcast jusqu'a 90 min) | MLX | 1.5B | EN/ZH |
 | [Qwen3.5-Chat](https://soniqo.audio/fr/guides/chat) | Texte → Texte (LLM) | MLX, CoreML | 0.8B | Multi |
+| [MADLAD-400](https://soniqo.audio/fr/guides/translate) | Texte → Texte (Traduction) | MLX | 3B | **400+** |
 | [PersonaPlex](https://soniqo.audio/fr/guides/respond) | Parole → Parole | MLX | 7B | EN |
 | [Silero VAD](https://soniqo.audio/fr/guides/vad) | Detection d'activite vocale | MLX, CoreML | 309K | Agnostique |
 | [Pyannote](https://soniqo.audio/fr/guides/diarize) | VAD + Diarisation | MLX | 1.5M | Agnostique |
@@ -139,6 +141,7 @@ Ensuite :
 ```bash
 audio transcribe recording.wav
 audio speak "Hello world"
+audio translate "Hello, how are you?" --to es
 audio respond --input question.wav --transcript
 audio-server --port 8080            # serveur HTTP / WebSocket local (OpenAI-compatible /v1/realtime)
 ```
@@ -166,6 +169,7 @@ import CosyVoiceTTS         // Synthese vocale avec clonage
 import KokoroTTS            // Synthese vocale (compatible iOS)
 import VibeVoiceTTS         // TTS long format / multi-locuteurs (EN/ZH)
 import Qwen3Chat            // Chat LLM embarque
+import MADLADTranslation    // Traduction multidirectionnelle entre 400+ langues
 import PersonaPlex          // Parole-a-parole full-duplex
 import SpeechVAD            // VAD + diarisation + empreintes
 import SpeechEnhancement    // Suppression de bruit
@@ -267,6 +271,16 @@ chat.chat(messages: [(.user, "Explain MLX in one sentence")]) { token, isFinal i
 }
 ```
 
+### Traduction — [guide complet →](https://soniqo.audio/fr/guides/translate)
+
+```swift
+import MADLADTranslation
+
+let translator = try await MADLADTranslator.fromPretrained()
+let es = try translator.translate("Hello, how are you?", to: "es")
+// → "Hola, ¿cómo estás?"
+```
+
 ### Detection d'activite vocale -- [guide complet →](https://soniqo.audio/fr/guides/vad)
 
 ```swift