Pocket TTS (Kyutai)

Overview

Pocket TTS is a lightweight, CPU-first text-to-speech system developed by Kyutai. It is designed for low-latency, real-time speech synthesis without requiring GPUs or external APIs, making it highly suitable for local, embedded, and modded environments like SkyrimNet.

Unlike cloud-based TTS solutions, Pocket TTS runs entirely locally, offering strong privacy guarantees and minimal setup overhead.

---

Pocket TTS is one of the most practical modern TTS solutions for SkyrimNet:

• Runs locally on CPU
• Streams audio in real-time
• Supports voice cloning
• Easy to integrate

It trades a bit of raw quality for speed, simplicity, and full offline capability

Key Features

• CPU-Only Inference

  • Runs efficiently on standard CPUs (no GPU required)
  • Uses ~2 CPU cores for real-time synthesis

• Low Latency Streaming

  • ~200ms to first audio chunk
  • Supports incremental audio generation (streaming pipeline)

• Fast Execution

  • Up to ~6× real-time speed on modern CPUs

• Small Model Size

  • ~100M parameters (extremely compact for TTS)

• Voice Cloning

  • Can replicate voices from short audio samples

• Multi-language Support

  • English, French, German, Portuguese, Italian, Spanish

• Unlimited Input Length

  • Handles arbitrarily long text inputs without chunking constraints

• Multiple Interfaces

  • Python API
  • CLI tool
  • Optional web server (FastAPI)

---

---

Strengths for SkyrimNet

Fully Local Execution

• No external API calls (ideal for offline mods)
• Avoids latency, cost, and privacy concerns

Real-Time NPC Dialogue

• Streaming output fits naturally with:
  • Dialogue systems
  • Interruptible speech
  • Reactive AI conversations

Lightweight Integration

• Simple install (just enable it on Easy settings)

Voice Personalization

• Voice cloning enables:
  • NPC-specific voices
  • Dynamic voice generation per character

---

Limitations

• Quality vs Size Tradeoff

  • Smaller model than SOTA cloud TTS (slightly less natural in some cases)

• Python-Centric

  • Requires bridging for C++ (Skyrim native integration)

• Voice Dataset Constraints

  • Voice cloning quality depends on input sample quality

• Ecosystem Maturity

  • New project, fewer tools compared to established TTS stacks

---

When to Use Pocket TTS in SkyrimNet

Best fit:

• Fully local AI setups
• Real-time NPC conversations
• Low-latency dialogue systems

Less ideal:

• Ultra-high-fidelity cinematic voice acting
• Complex emotional prosody (compared to large cloud models)

---

Comparison Between TTS

System	Type	Runs Local	Latency	Quality	Voice Cloning	Resource Usage	Best Use Case
Piper	Lightweight VITS	✅ CPU	⚡ Extremely low	❌ Basic	❌ No	🟢 Very low (runs on weak CPUs)	Fast NPC chatter
XTTS v2	Neural (Coqui)	✅ GPU	🐢 High	✅ High	✅ Yes	🟡 Moderate (GPU required,aprox 3GB VRAM)	High-quality voices
Inworld TTS	Large-scale AI	❌ (API / heavy local)	⚡ Low–Medium	🧠 SOTA / expressive	✅ Yes	🟢 None (but costs $) External service	AAA dialogue systems
Pocket TTS	Flow-based (Kyutai)	✅ CPU	⚡ Very low	✅ Good+	✅ Yes	🟡 Moderate (multi-core CPU)	Real-time local AI