Coqui TTS Latency by GPU (Full Benchmark)

Test Methodology

We benchmarked Coqui XTTS-v2 across six GPUs on GigaGPU dedicated servers. XTTS-v2 is the leading open-source text-to-speech model with voice cloning capabilities. Tests generate a standardised 30-second speech clip (~75 words) and measure end-to-end latency including model inference and vocoder pass. GigaGPU’s Coqui TTS hosting provides pre-configured servers for production TTS deployment. For interactive data, visit our TTS latency benchmarks tool.

Latency Benchmarks by GPU

Lower latency means faster audio generation. For real-time voice applications, the audio must be generated faster than playback speed (RTF below 1.0).

Coqui XTTS-v2 (30-second clip, single request)

All GPUs achieve sub-1.0 RTF, meaning every card can generate speech faster than real-time playback. The RTX 5090 generates 30 seconds of speech in 2.1 seconds (14x faster than real-time). The RTX 3050 is the slowest at 0.85 RTF but still faster than real-time. For comparison with other TTS models, see our best GPU for TTS voice AI guide.

Real-Time Factor Comparison

Real-time factor (RTF) is the ratio of generation time to audio duration. RTF below 1.0 means the GPU generates speech faster than playback. Lower is better.

RTF by Audio Length

RTF improves slightly with longer clips due to reduced per-request overhead. The RTX 3050 approaches its real-time limit at 10-second clips (0.92 RTF), leaving minimal headroom for streaming applications.

Voice Cloning Performance

XTTS-v2 supports zero-shot voice cloning from a short reference audio clip. We measured the overhead of voice cloning compared to using a default voice.

Voice cloning adds 14-18% latency overhead. On the RTX 3050, cloned voice generation approaches the real-time boundary (0.99 RTF). For reliable voice cloning, use an RTX 4060 or better. For full voice agent pipelines that combine Whisper + LLM + TTS, see our multi-model GPU guide and speech model hosting.

Cost per Audio Hour

We calculated the cost to generate one hour of speech audio using XTTS-v2 at sustained throughput.

Self-hosted Coqui TTS costs $0.08-$0.13 per audio hour, which is 100-275x cheaper than ElevenLabs API pricing. Even the cheapest GPU generates one hour of audio for under 9 cents. See our GPU vs API cost analysis for general cost comparison methodology.

Concurrent Stream Performance

For applications serving multiple users, concurrent TTS generation is critical. We measured RTF with multiple simultaneous generation requests on the RTX 3090.

The RTX 3090 handles up to 4 concurrent TTS streams while maintaining real-time generation. Beyond 4 streams, RTF exceeds 1.0 and users experience buffering. For higher concurrency, use the RTX 5090 (supports 8+ streams) or multi-GPU clusters.

GPU Recommendations for Coqui TTS

Best overall: RTX 3090. At 0.18 RTF and $0.081 per audio hour, the RTX 3090 is the cost-efficiency champion for TTS. It handles 4 concurrent streams and voice cloning comfortably. The 24 GB VRAM leaves room for running an LLM alongside TTS for conversational AI.

Best for real-time voice applications: RTX 5090. At 0.07 RTF, the 5090 generates speech almost instantly and handles 8+ concurrent streams. Essential for low-latency voice agents and real-time conversational AI. Pairs well with Whisper for speech-to-text.

Best budget: RTX 4060. At 0.39 RTF and $0.078 per audio hour, the RTX 4060 delivers the lowest cost per audio hour. Handles single-user TTS and basic voice cloning. Good for development, prototyping, and batch audio generation.

Best mid-range: RTX 5080. At 0.12 RTF the 5080 handles voice cloning with headroom and supports 4-5 concurrent streams. Good balance of speed and cost for production TTS.

For full voice agent pipelines, see our TTS voice AI GPU guide, AI agents GPU guide, and RAG pipeline GPU guide. Related benchmarks: LLaMA 3 8B benchmark, Mistral 7B benchmark, Stable Diffusion benchmark, and Whisper RTF benchmark.