MOSS-Speech: Towards True Speech-to-Speech Models Without Text Guidance

MOSS-Speech represents a research effort to move away from 'cascaded' pipelines (ASR -> LLM -> TTS) toward 'native' speech models. While academically relevant, the project has minimal defensibility (0 stars, though 23 forks indicate some researcher interest). It enters a hyper-competitive space dominated by frontier labs: OpenAI's GPT-4o and Google's Gemini Live already implement native multimodal architectures that achieve these goals with far superior data and compute resources. Within the open-source ecosystem, Kyutai's 'Moshi' has already established itself as the leading reference for low-latency, end-to-end speech-to-speech. MOSS-Speech's specific angle—omitting text guidance entirely—is a technical nuance that may help with paralinguistic cues but faces an uphill battle against models that use text as a semantic stabilizer. The displacement horizon is near-immediate as larger, more robust models are already in production.