From a95500ef675252f539039017f23a8074c01d3fb3 Mon Sep 17 00:00:00 2001
From: Hermes <hermes@niklashmotion.art>
Date: Sun, 7 Jun 2026 13:10:51 +0200
Subject: [PATCH] Voice MCP Server: STT (Moonshine Tiny) + TTS (KittenTTS Nano)

---
 Dockerfile      |  25 +++++++++
 voice_server.py | 138 ++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 163 insertions(+)
 create mode 100644 Dockerfile
 create mode 100644 voice_server.py

diff --git a/Dockerfile b/Dockerfile
new file mode 100644
index 0000000..a81c174
--- /dev/null
+++ b/Dockerfile
@@ -0,0 +1,25 @@
+FROM python:3.12-slim
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    libsndfile1 \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+# Install CPU-only torch first to prevent CUDA download
+RUN pip install --no-cache-dir \
+    --extra-index-url https://download.pytorch.org/whl/cpu \
+    torch==2.6.0 \
+    && pip install --no-cache-dir mcp httpx
+
+# Install Moonshine (STT)
+RUN pip install --no-cache-dir moonshine-voice
+
+# Install KittenTTS (TTS) — CPU-only
+RUN pip install --no-cache-dir \
+    https://github.com/KittenML/KittenTTS/releases/download/0.8.1/kittentts-0.8.1-py3-none-any.whl
+
+COPY voice_server.py .
+
+EXPOSE 8000
+CMD ["python", "voice_server.py"]
diff --git a/voice_server.py b/voice_server.py
new file mode 100644
index 0000000..bde82f2
--- /dev/null
+++ b/voice_server.py
@@ -0,0 +1,138 @@
+#!/usr/bin/env python3
+"""Voice MCP Server — STT via Moonshine, TTS via KittenTTS.
+
+HTTP transport for Hermes MCP client. Tools:
+  voice_stt(file_path, language) → transcription text
+  voice_tts(text, voice, speed)  → path to generated .wav
+  voice_pipeline(audio_path)     → full STT → TTS roundtrip
+"""
+
+import hashlib
+import os
+import tempfile
+from pathlib import Path
+
+from mcp.server.fastmcp import FastMCP
+
+app = FastMCP("voice")
+
+# ── Lazy model loading (RAM-friendly) ────────────────────
+
+_stt = None
+_tts = None
+_model_dir = Path(os.environ.get("MODEL_DIR", "/models"))
+
+
+def get_stt():
+    global _stt
+    if _stt is None:
+        from moonshine_voice import Transcriber, ModelArch
+        from moonshine_voice.download import get_model_for_language
+        model_path, model_arch = get_model_for_language(
+            "en", ModelArch.TINY, cache_root=str(_model_dir)
+        )
+        _stt = Transcriber(model_path=model_path, model_arch=model_arch)
+    return _stt
+
+
+def get_tts():
+    global _tts
+    if _tts is None:
+        from kittentts import KittenTTS
+        _tts = KittenTTS(
+            "KittenML/kitten-tts-nano-0.8-int8",
+            cache_dir=str(_model_dir),
+        )
+    return _tts
+
+
+# ── Tools ────────────────────────────────────────────────
+
+@app.tool()
+async def voice_stt(file_path: str, language: str = "auto") -> str:
+    """Transcribe speech audio to text using Moonshine.
+
+    Args:
+        file_path: Absolute path to audio file (.wav)
+        language: Language code or 'auto' (default: auto)
+    Returns:
+        Transcribed text
+    """
+    from moonshine_voice.utils import load_wav_file
+    stt = get_stt()
+    audio_data, sample_rate = load_wav_file(file_path)
+    result = stt.transcribe_without_streaming(audio_data, sample_rate)
+    return " ".join(l.text for l in result.lines)
+
+
+@app.tool()
+async def voice_tts(
+    text: str,
+    voice: str = "Jasper",
+    speed: float = 1.0,
+) -> str:
+    """Convert text to speech using KittenTTS.
+
+    Args:
+        text: Text to speak
+        voice: One of Bella, Jasper, Luna, Bruno, Rosie, Hugo, Kiki, Leo
+        speed: Speech speed (0.5-2.0, default 1.0)
+    Returns:
+        Absolute path to generated .wav file
+    """
+    tts = get_tts()
+
+    # Stable filename based on content hash
+    h = hashlib.sha256(f"{text}_{voice}_{speed}".encode()).hexdigest()[:12]
+    out_path = os.path.join(tempfile.gettempdir(), f"tts_{h}.wav")
+
+    tts.generate_to_file(text, out_path, voice=voice, speed=speed)
+    return out_path
+
+
+@app.tool()
+async def voice_pipeline(audio_path: str, tts_voice: str = "Jasper") -> str:
+    """Full voice pipeline: STT → echo back as TTS.
+    Useful for testing the roundtrip.
+
+    Args:
+        audio_path: Path to input audio file
+        tts_voice: Voice for TTS response
+    Returns:
+        JSON with transcribed text and path to response audio
+    """
+    import json
+
+    # STT
+    from moonshine_voice.utils import load_wav_file
+    stt = get_stt()
+    audio_data, sample_rate = load_wav_file(audio_path)
+    result = stt.transcribe_without_streaming(audio_data, sample_rate)
+    text = " ".join(l.text for l in result.lines)
+
+    # TTS
+    tts = get_tts()
+    h = hashlib.sha256(text.encode()).hexdigest()[:12]
+    out_path = os.path.join(tempfile.gettempdir(), f"tts_{h}.wav")
+    tts.generate_to_file(text, out_path, voice=tts_voice)
+
+    return json.dumps({
+        "transcription": text,
+        "response_audio": out_path,
+    })
+
+
+@app.tool()
+async def voice_available_voices() -> str:
+    """List available TTS voices."""
+    tts = get_tts()
+    return str(tts.available_voices)
+
+
+# ── Entry Point ──────────────────────────────────────────
+
+if __name__ == "__main__":
+    import sys, uvicorn
+    port = int(os.environ.get("PORT", "8000"))
+    print(f"Voice MCP starting on port {port}...", file=sys.stderr)
+    uvicorn.run(app.streamable_http_app(), host="0.0.0.0", port=port)