BlueRose Labs fine-tuned Kyutai's Pocket TTS 18M into Silencio-2, a CC-BY-4.0 English voice model that takes inline tags like [sigh], [laugh] and whisper cues, clones a voice from a short clip, and runs about 20x faster than real time on a plain CPU.
ElevenLabs released SFX v3, its new closed text-to-sound-effects model. It follows multi-sound prompts more closely, handles music and vocals better, and is live in ElevenCreative and the API as eleven_text_to_sound_v3, in limited access.
Cartesia launched Managed Agents, a closed end-to-end voice agent platform built around its own Sonic and Ink models. It ranks #1 of 30 on Artificial Analysis's Speech Agent Arena and costs $0.03 per call minute plus LLM and telephony.
Inworld added Reasoning to Realtime TTS-2: send a line plus context, and the model writes its own delivery direction before speaking it, returned in the API response. Opt-in, same model ID, same $25 per 1M characters.
Edge0 open-sourced Audio8-TTS-Realistic, a 0.6B Apache-2.0 text-to-speech model that clones a voice from one short clip in 23 languages, with a free Hugging Face playground and an OpenAI-compatible server.
Phonic launched Phonic-1, a closed speech-to-speech voice model that debuted at #1 on Artificial Analysis' Speech to Speech Index (87.1), answers in about 0.61s, speaks 50+ languages and runs tool calls mid-sentence. It's live on the Phonic API at $0.07 a minute and in LiveKit Agents.
Deepgram is acquiring Daily, the team behind the open-source Pipecat voice agent framework, which stays BSD-2 licensed under a new Pipecat Foundation. It also launched RCS 0.1, an open alpha standard that carries tone and delivery tags between STT, LLM and TTS.
Suno launched an official MCP server at api.suno.com/mcp. Connect it to Claude, ChatGPT, Gemini, Claude Code or Cursor with a Suno sign-in, no API key, and generate, remix and extend songs on your normal credits.
Rime launched Prism, a closed speech-to-speech model that listens to raw audio, writes an inspectable text reply, then speaks it: 0.57 s to first audio, 56.6 on τ-Voice, $0.08 a minute and a free $50 monthly tier.
AssemblyAI launched AssemblyAI-TTS-1, its first text-to-speech model: 90 ms to first audio in production, 4,000+ voices, instant cloning, six languages and a flat $15 per 1M characters with unlimited concurrency.
NVIDIA released Nemotron 3.5 Transcribe, an open-weight 30B mixture-of-experts speech model with about 4B active parameters that transcribes 40 languages, translates 31, labels up to eight speakers and handles hour-long recordings on one H100.
Gnani AI launched Prisma v3.0, a speech-to-text model for nine Indian languages plus English with auto language detection, 200 keyterms per request and four output modes, scoring 16.93 macro WER on Vistaar against 19.98 for ElevenLabs Scribe v2.
Liquid AI released LFM3-Audio-2B, an open-weight 2B audio model that listens, talks back in 24 kHz speech, transcribes, translates and turns spoken requests into tool calls across 18 languages, with GGUF, MLX and ONNX builds for laptops and edge devices.
Fish Audio's new transcribe-stream-1 streams live transcripts over WebSocket and decides when a speaker has actually finished, tells a quick "mm-hm" from a real interruption, and costs $0.36 per hour in preview.
Maya Research released the weights for Maya 2 Native, a 2.5B streaming text-to-speech model for 11 Indic languages and English with 20 built-in voices, cross-lingual voice cloning from a few seconds of audio, and an Apache-2.0 license.
Tencent Hunyuan released Hy-SimulMT, an Apache-2.0 30B-A3B model that translates Chinese and English while the speaker is still talking, with about 3 seconds of streaming lag and optional voice cloning through Qwen3-TTS.
Ollama v0.40.11 adds speech generation. Pull Gradium's tiny Apache-2.0 Phonon voice, pipe any local chat model into it, and stream audio from an OpenAI-compatible /v1/audio/speech endpoint on your own machine.
Tesslate just dropped Lure, an Apache-2.0 English TTS model that clones a voice from a few seconds of audio and actually performs (laugh), (sigh) and (sharp inhale) tags. It's only 0.6B parameters, there's a faster Turbo version, and it runs on a Mac, an NVIDIA GPU or a plain CPU.
Fermion Research just open-sourced Phonon-2.5, a speech-to-text model for 25 languages that squeezes NVIDIA's 2.5 GB Parakeet v3 into 156 MB and still beats it on average word error. It runs on Apple silicon or a plain CPU, gives you word timestamps, and can serve an OpenAI-style transcription endpoint on your own machine.
Block's small AI lab just showed Dessa, a speech-to-speech model trained from scratch on recorded conversation with no text and no pretrained LLM underneath. It listens while it talks, laughs mid-sentence, and runs locally on a MacBook Pro M4 Max. Right now it's a research preview for AI safety researchers only.
ElevenCreative's new Chimera template pairs Eleven v4 voices with Seedance 2.5 video, so you can direct a character's read with tags like [whispers] and [laughs] and get the face to act it out. Here's what shipped, what it can't do yet, and who can use it.
ComfyUI v0.42.4 quietly added two Mirelo partner nodes: Video to SFX, which times Foley to your clip and keeps the dialogue, and Text to SFX, which makes up to 10 minutes of sound from a prompt. Both run sfx-2.0 at roughly $0.0133 per second, with official templates already in the library.
HeyGen just released Voice Turbo, a faster and cheaper tier of its in-house TTS model. It now leads the Artificial Analysis Controlled Voice Arena at 1,202 Elo, lists at $4 per 1M characters, and a heygen-voice-1-turbo-free endpoint costs nothing through November 30. It's also live on fal and in ComfyUI.
ElevenLabs just launched Eleven Realtime, two speech-to-speech models that fold listening, turn-taking, reasoning and voice into one agent. v1 Max scores 98.8% on Big Bench Audio with 0.57 s to first audio, both work with 17,000+ library voices in 90+ languages, and the OpenAI-compatible Realtime API starts at $0.08 a minute plus LLM tokens.