Bark vs XTTS-v2 vs Kokoro: TTS Model Selection

Self-hosted text-to-speech has three serious contenders: Bark for expressive generation with sound effects, XTTS-v2 for multilingual voice cloning, and Kokoro for fast, lightweight synthesis. Each model occupies a different niche on dedicated GPU hardware, and picking the wrong one wastes either quality or compute.

Model Overview

The Real-Time Factor (RTF) is the critical production metric. An RTF under 1.0 means audio generates faster than playback speed. Bark barely crosses this line at 0.8, making it impractical for live applications. XTTS-v2 at 0.15 generates audio nearly 7x faster than playback. Kokoro at 0.03 is roughly 33x faster — fast enough for real-time voice agents with negligible latency.

Voice Quality Comparison

Bark produces the most expressive output. It can generate laughter, gasps, pauses, and even background music through text annotations. Voice quality is natural but inconsistent — the same prompt can produce different prosody patterns across generations. Best suited for creative content, audiobooks, and media production where expressiveness matters more than consistency.

XTTS-v2 delivers the most versatile voice cloning. A 6-second audio clip is enough to capture a speaker’s characteristics and reproduce them across languages. Quality is consistent and production-ready. The ideal choice for personalised voice interfaces and multilingual applications.

Kokoro prioritises speed and simplicity. Voice quality is clean and natural for English but lacks the expressive range of Bark or the cloning flexibility of XTTS-v2. Pre-built voice presets cover common use cases without requiring any audio reference files.

GPU Requirements

All three models are lightweight enough to share a GPU with an LLM. Kokoro at 0.5 GB VRAM is particularly suited to voice agent pipelines where Whisper, an LLM, and TTS all run on a single card. XTTS-v2 and Bark require more VRAM headroom but still fit alongside 7B-class models on an RTX 3090.

Use Case Matching

Voice agents and phone bots: Kokoro or XTTS-v2. Latency is the priority. Bark is too slow for real-time conversation. XTTS-v2 adds voice cloning if you need a branded voice; Kokoro is simpler if you just need clear, fast speech.

Audiobook and podcast generation: Bark. The expressive range produces natural-sounding narration with appropriate emotional inflection. Batch processing means the slower speed is acceptable. Pair with Whisper for a full audio pipeline.

Multilingual applications: XTTS-v2. It is the only option that handles 17 languages with voice cloning across all of them. A single reference clip can generate speech in any supported language while maintaining the speaker’s voice characteristics.

High-volume notification systems: Kokoro. Generating thousands of short audio clips per minute for automated calls, alerts, or announcements requires Kokoro’s throughput advantage.

Deployment Notes

Wrap any of these models behind a FastAPI server for production use. XTTS-v2 and Kokoro support streaming output, which means the first audio chunk can play while the rest generates — critical for real-time translation pipelines.

For choosing between these models alongside other pipeline decisions, see image model selection and embedding model comparison. The best GPU for inference guide covers hardware selection, and benchmark data is available for throughput planning.