Show HN: Inworld TTS – high-quality, affordable, and low-latency TTS
Hi HN, Igor here, one of the engineers behind this project. High-quality voice APIs are usually either expensive, slow, or both. Cheaper and faster solutions very often lack realism. We decided to build Inworld TTS to bridge this gap. We just released two multilingual models. Our small model, named TTS-1, is on par with SOTA models quality-wise given objective metrics WER/SIM/DNSMOS. A larger model, TTS-1-Max, is even better. It can produce more nuanced speech and has ~3.5% better WER across all 11 supported languages averaged. Both models also support markup tags (e.g. prepend "[happy]" to the text to make the generation more enthusiastic, etc). The models are built with LLaMA 1B and 8B being the SpeechLM backbones for TTS-1 and TTS-1-Max respectively. We up-trained both models on a mixture of text and audio, then finetuned on text-audio pairs and polished final chec
Why it failed
Hi HN, Igor here, one of the engineers behind this project. High-quality voice APIs are usually either expensive, slow, or both. Cheaper and faster solutions very often lack realism. We decided to build Inworld TTS to bridge this gap. We just released two multilingual models. Our
- Kind
- startup
- Status
- Shut down
- Inventor
- rogilop
- Ended
- 2025
- Source
- hackernews