Coda is a generative text-to-speech (TTS) model developed by Rime as the successor to its previous flagship, Arcana. The model is built on an architecture that integrates a sophisticated Large Language Model (LLM) backbone with a dedicated speech inference engine. It is designed to produce high-fidelity audio characterized by enhanced naturalness, improved prosody, and a reduction in synthesized artifacts compared to earlier iterations.
A distinguishing feature of Coda is its training methodology, which utilizes full-duplex conversational data. Rather than relying solely on studio narrations or audiobooks, the model was trained on a proprietary dataset of real human-to-human dialogue. This approach allows the model to better replicate the organic flow, interruptions, and contextual nuances found in live speech, making it suitable for applications requiring high emotional depth and conversational realism.
The model provides native multilingual support across a variety of languages, including English, Spanish, French, Portuguese, German, and Japanese, with further expansions into Arabic and Hindi. It is optimized for speed, maintaining sub-100ms model latency to support real-time applications. Additional technical capabilities include the generation of word-level timestamps for precise audio-text alignment and a specialized spell() function for handling the articulation of characters and numeric sequences.
Coda provides access to a library of over 180 unique voices, reflecting a diverse range of demographics and speaking styles. The model's design prioritizes stability and consistency across its voice lineup, allowing for shared voice identities to be maintained across multiple supported languages.