Hosted text-to-speech models

Text-to-speech (TTS) models convert written text into natural-sounding speech, solving problems such as generating voiceovers for videos, enabling screen readers for accessibility, powering interactive voice response systems, and providing real-time narration in navigation or e-learning applications. In production, TTS is typically integrated via API calls from applications that need to stream audio on demand—common architectures use queuing for batch jobs or low-latency streaming for conversational agents.

Choosing among the 37 models being onboarded (including coqui/XTTS-v2, multiple Qwen/Qwen3-TTS-12Hz variants, OpenMOSS-Team/MOSS-TTS, k2-fsa/OmniVoice, SWivid/F5-TTS, ai4bharat/indic-parler-tts, and Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign) involves a trade-off between quality, speed, and model size. Larger models (e.g., 1.7B parameters) generally produce richer, more expressive speech at the cost of higher latency and compute; smaller models (e.g., 0.6B) offer faster inference and lower cost, suitable for high-throughput or latency-sensitive applications.

Using a hosted API eliminates the overhead of provisioning GPU infrastructure, managing model updates, and scaling for variable demand—making it more economical than self-hosting for most call volumes below tens of thousands of requests per minute.