Voice, speech, and meeting AI

A compact 2B audio-to-text scope with 14-language coverage centres multilingual transcription rather than voice generation or meeting-agent features.

Training and inference work on difficult real-world recordings focus on transcription under noise and uncontrolled acoustic conditions.

Coverage of Mandarin, English, Chinese dialects, code-switching, songs, and multiple speakers addresses ASR shaped by language mixing and speaker complexity.

Speech, sound, music, captioning, time-aware questions, and ASR give MOSS-Audio a broader audio-understanding scope beyond transcription alone.

A 600M streaming design with multilingual locales and latency and throughput tables exposes the balance between performance and language coverage in live transcription.

An on-device CLI focused on fast Whisper transcription prioritises local processing speed over a complete voice application.

Detailed prosody and emotional-delivery controls support expressive direction beyond simply converting text to speech.

A very small footprint targets voice generation under lightweight deployment constraints.

With 82M parameters, voice materials, samples, and an inference library, this model offers a compact path from evaluation to implementation.

Voice design, dialogue, realtime speech, compact generation, and sound effects place the MOSS-TTS family across several audio-output jobs.

A tiny multilingual model with CPU-friendly realtime positioning targets responsive local speech on modest hardware.

Japanese-centred generation, English support, style transfer, and a zero-shot voice path cover Japanese delivery and voice adaptation.

Local ONNX inference, 31-language coverage, and expression tags show multilingual speech with controllable delivery across device types.

Multilingual generation, voice design, controllable cloning, and streaming support combine tailored voices with live TTS output.

A local-first studio combines cloning, synthesis, effects, and app workflows for work that needs an end-user production environment rather than only a model.

WebRTC rooms, telephony, tools, and deployment paths show voice agents joining live calls and interacting during the session.

Audio and video pipeline stages, transports, client SDKs, and structured flows show realtime conversation components being assembled and swapped.

Desktop meeting notes, dictation, projects, and local app state carry spoken work into organised desktop tasks rather than ending at transcription.

Live transcription with local or external summary choices makes a meeting assistant’s processing boundary explicit.

Conversational note-taking distinguishes a responsive meeting companion from a passive recording pipeline.

A full-duplex speech-to-speech design with role prompts and voice conditioning addresses interruption handling and conversational persona in live agents.