feat(vllm): add TTS audio generation endpoint via vLLM Omni

Add POST /v1/audio/speech endpoint supporting text-to-speech with
vLLM Omni, following the OpenAI audio API convention. Includes Rust
protocol types, stream aggregation, model discovery for audio-capable
workers, and a Python handler that builds TTS engine inputs with
prompt length estimation and WAV encoding.

Signed-off-by: Hatem Elseidy <hatem.elseidy@epicgames.com>

Author: Hatem Elseidy

components/src/dynamo/common/protocols/__init__.py

@@ -5,15 +5,22 @@
 
 This module provides protocol types for various modalities:
 - video_protocol: NvCreateVideoRequest, NvVideosResponse for video generation
+- audio_protocol: NvCreateAudioSpeechRequest, NvAudiosResponse for audio generation
 """
 
+from dynamo.common.protocols.audio_protocol import (
+    NvAudiosResponse,
+    NvCreateAudioSpeechRequest,
+)
 from dynamo.common.protocols.video_protocol import (
     NvCreateVideoRequest,
     NvVideosResponse,
     VideoData,
 )
 
 __all__ = [
+    "NvAudiosResponse",
+    "NvCreateAudioSpeechRequest",
     "NvCreateVideoRequest",
     "NvVideosResponse",
     "VideoData",

components/src/dynamo/common/protocols/audio_protocol.py

@@ -0,0 +1,92 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Protocol types for audio speech generation.
+
+These types match the Rust protocol types in lib/llm/src/protocols/openai/audios.rs
+to ensure compatibility with the Dynamo HTTP frontend.
+"""
+# TODO: Replace these Pydantic models with Python bindings to the Rust protocol types once PyO3 bindings are available.
+
+from typing import Optional
+
+from pydantic import BaseModel
+
+
+class AudioNvExt(BaseModel):
+    """NVIDIA extensions for audio speech generation requests.
+
+    Matches Rust NvExt in lib/llm/src/protocols/openai/audios/nvext.rs.
+    """
+
+    annotations: Optional[list[str]] = None
+    """Annotations for SSE stream events."""
+
+    task_type: Optional[str] = None
+    """Task type (e.g. 'tts', 'voice_clone')."""
+
+    language: Optional[str] = None
+    """Language code (e.g. 'en', 'zh')."""
+
+    instructions: Optional[str] = None
+    """Additional instructions for speech generation."""
+
+    ref_audio: Optional[str] = None
+    """Base64-encoded reference audio for voice cloning."""
+
+    ref_text: Optional[str] = None
+    """Reference text corresponding to ref_audio."""
+
+    max_new_tokens: Optional[int] = None
+    """Maximum number of tokens to generate."""
+
+    seed: Optional[int] = None
+    """Random seed for reproducibility."""
+
+
+class NvCreateAudioSpeechRequest(BaseModel):
+    """Request for audio speech generation (/v1/audio/speech endpoint).
+
+    Matches Rust NvCreateAudioSpeechRequest in lib/llm/src/protocols/openai/audios.rs.
+    """
+
+    # Required fields
+    input: str
+    """The text to generate audio for."""
+
+    model: str
+    """The model to use for audio generation."""
+
+    voice: str
+    """The voice to use for generation."""
+
+    # Optional fields
+    response_format: Optional[str] = None
+    """Audio format: mp3, wav, opus, aac, flac, pcm."""
+
+    speed: Optional[float] = None
+    """Playback speed (0.25 to 4.0, default 1.0)."""
+
+    nvext: Optional[AudioNvExt] = None
+    """NVIDIA extensions."""
+
+
+class NvAudiosResponse(BaseModel):
+    """Response structure for audio speech generation.
+
+    Matches Rust NvAudiosResponse in lib/llm/src/protocols/openai/audios.rs.
+    Internal transport uses base64-encoded audio. The HTTP handler decodes
+    this to return raw binary audio to clients.
+    """
+
+    audio_b64: str
+    """Base64-encoded audio bytes."""
+
+    content_type: str
+    """MIME content type (e.g. 'audio/mpeg', 'audio/wav')."""
+
+    model: str
+    """Model used for generation."""
+
+    created: int
+    """Unix timestamp of creation."""

components/src/dynamo/common/utils/output_modalities.py

@@ -6,6 +6,7 @@
 
 from pydantic import BaseModel
 
+from dynamo.common.protocols.audio_protocol import NvCreateAudioSpeechRequest
 from dynamo.common.protocols.image_protocol import NvCreateImageRequest
 from dynamo.common.protocols.video_protocol import NvCreateVideoRequest
 from dynamo.llm import ModelType
@@ -91,8 +92,7 @@ def parse_request_type(
         return NvCreateVideoRequest(**raw_request), RequestType.VIDEO_GENERATION
 
     if modality is OutputModality.AUDIO:
-        # Audio protocol types are not yet defined; pass through the raw dict.
-        return raw_request, RequestType.AUDIO_GENERATION
+        return NvCreateAudioSpeechRequest(**raw_request), RequestType.AUDIO_GENERATION
 
     # Text Modality
     return raw_request, RequestType.CHAT_COMPLETION