About NVIDIA TTS NIM Microservice#
The NVIDIA Text-to-Speech (TTS) NIM microservice synthesizes natural-sounding speech from text. It packages pre-trained NeMo models with the full NVIDIA inference stack into self-contained containers that handle model download, optimization, and serving.
The NVIDIA TTS NIM microservice supports two synthesis modes:
Offline: Generates the complete audio and returns it in a single response. Simpler to use, but subject to a 4 MB gRPC message size limit.
Streaming: Returns audio in chunks as they are generated. Provides lower time-to-first-audio and handles arbitrarily long text.
Available Models#
Model |
Languages |
Modes |
Key Capability |
|---|---|---|---|
Magpie TTS Multilingual |
English, Spanish, French, German, Mandarin, Vietnamese, Italian, Hindi, Japanese, Korean, Arabic, Portuguese |
Streaming + Offline |
Multi-voice, multi-language synthesis |
Magpie TTS Zeroshot |
English, Spanish, French, German, Mandarin, Vietnamese, Italian, Hindi, Japanese, Arabic, Portuguese, Korean |
Streaming + Offline |
Multilingual voice cloning from a reference audio sample |
Chatterbox TTS Multilingual |
23 languages incl. Arabic, English, Spanish, French, German, Hindi, Japanese, Korean, Mandarin, Portuguese, Russian |
Streaming + Offline |
Emotion exaggeration |
Note
Magpie TTS Zeroshot requires access approval. Request access through this form.
For GPU requirements and all model profiles, refer to the TTS support matrix.
Key Capabilities#
Voice Selection and Emotional Styles#
Each TTS model includes multiple voices organized by language and locale. The Magpie Multilingual model provides voices with optional emotional style variants, such as Magpie-Multilingual.EN-US.Aria.Angry and Magpie-Multilingual.FR-FR.Pascal.Calm. Refer to Voices and Emotional Styles for the naming convention and available emotions.
Voice Cloning#
The Magpie TTS Zeroshot model clones a voice from a short reference audio recording (3–10 seconds, 16-bit mono WAV at 22.05 kHz or higher). Refer to Voice Cloning for audio prompt preparation, quality tuning, and examples.
Batch Synthesis#
The WebSocket realtime client supports synthesizing speech from text files and processing multiple lines in parallel for higher throughput. Refer to Batch Synthesis for details.
SSML Customization#
The TTS NIM microservice supports a subset of Speech Synthesis Markup Language (SSML) for controlling phoneme pronunciation. Refer to Customizing TTS Models for details and examples.
Custom Pronunciation#
Define a text-based dictionary that maps words to IPA phonetic representations. The TTS model applies these custom pronunciations during synthesis. Refer to Phoneme Support for supported phonemes.
Next Steps#
TTS Tutorial: Deploy your first TTS NIM and synthesize speech step by step.
Deploy and Run TTS NIM: Docker and Helm deployment with inference examples for all models and APIs.
Voices and Emotional Styles: Voice naming convention, speakers, and emotional variants.
Voice Cloning: Clone a voice from a reference audio recording.
Batch Synthesis: Synthesize from text files with parallel processing.
Customizing TTS Models: SSML tags and custom pronunciation dictionaries.
Deploying Custom Models as NIM: Convert, build, and deploy a custom NeMo TTS model as a NIM.
Phoneme Support: Supported IPA phonemes for English.
TTS Support Matrix: GPU requirements and model profiles.