Murf AI just released Falcon 2, a text-to-speech model priced at $0.01 per minute—one-fifth the cost of ElevenLabs. It clocks response times under 100 milliseconds and scored ahead of OpenAI's Realtime API on Artificial Analysis benchmarks. Supporting over 35 languages and 10,000 concurrent calls, it aims to slash the cost of running real-time enterprise voice agents.
Adobe just made its Firefly audio suite generally available, letting creators generate commercially safe music, voiceovers, and sound effects directly in their workflow. The update also integrates Google's Gemini Omni Flash model and rolls out a free tier with daily generations for Firefly AI Assistant. Why it matters: Content creators can now bypass licensing nightmares and voice recording logistics inside a single platform.
Meta AI just launched its official macOS desktop app, headlined by a system-wide voice dictation tool. Users can simply hold down the 'fn' key and start talking to transcribe text into any app with ultra-high accuracy, eliminating the need to switch windows for hands-free typing.
ElevenLabs just made its Eleven v3 Conversational speech model generally available via ElevenAgents and its API. The real-time model covers 70+ languages and 11,000+ voices, adding audio tags that let developers cue emotions like whispering, laughing, or sarcasm. Starting at $0.05 per 1,000 characters, it brings humanlike nuance to voice agents.
Audio8 just dropped Audio8TS, an ultra-compact text-to-speech model with roughly 100M parameters and a tiny 1.7GB footprint. It pulls off zero-shot voice cloning from just a few seconds of reference audio and even ships with an ONNX INT8 build for low-resource CPU inference. Running high-quality voice cloning locally on a basic laptop is now genuinely practical.
Researchers just open-sourced VoiceMem, a dedicated long-term memory system built specifically for real-time voice agents. It splits memory into a left brain for facts and a right brain for emotion, speculatively fetching memories while the user is still speaking with just 134 ms latency. Why it matters: It keeps voice companions responsive without freezing to dig through conversation logs, cutting per-turn overhead to roughly 430 tokens.
Pika just launched Pika Audio, a family of four foundation models covering video soundtracks, music, sound effects, and text-to-speech. The company claims its training and inference stack makes them up to 20x cheaper than rival models, with speech generation 9x cheaper than ElevenLabs v3. They are available exclusively through the Pika API Club.
Grok Bot just rolled out phone calling for paid users, letting the AI sit on hold, book appointments, and negotiate bills on U.S. numbers. You can follow a live transcript and type replies mid-call if the bot gets stuck, but you are legally responsible for every call it makes.
Xiaomi researchers just open-sourced MiDashengLM-Gen, an end-to-end framework capable of generating coherent audio scenes blending speech, background music, sound effects, and ambient noise in a single pass. Powered by a fine-tuned Qwen3-1.7B and Flow Matching DiT, it turns structured text prompts into unified 16 kHz audio under an Apache 2.0 license.
Deepgram just launched Flux TTS, a text-to-speech model built specifically for live conversations. It delivers latency as low as 80ms and handles interruptions, order changes, and rapid numbers mid-call. Instead of flashy demos, it is built for real production deployments, and it is free for developers until September 12th.
IndexTeam just released IndexTTS-2.5, upgrading its open-source zero-shot speech synthesis model. The new release adds support for Chinese, English, Japanese, Spanish, and Arabic, featuring finer emotion manipulation, precise phoneme pronunciation overrides, and vLLM production serving.
ElevenLabs just rolled out its Dubbing v2 API, letting developers build AI dubbing directly into their apps and workflows. The updated model anchors to the original audio performance to preserve tone, delivery, and emotion across 90+ languages. It automatically syncs timing and handles background music and multi-speaker audio with sharper regional accent support.
xAI just upgraded its Imagine Video 1.5 model, adding image and voice references alongside native 1080p text-to-video generation. Creators can now submit up to seven reference images to lock faces, products, or scenes across multiple clips, keeping character voices consistent throughout. Native 1080p and text-to-video are live for everyone, while reference tools roll out first to US paid subscribers.
Avi Schiffmann just launched Friend 2.0, giving the loneliness-fighting AI necklace a built-in speaker for real-time voice conversations. But the talkative pendant now costs $249, up steeply from its original $99 price tag. Why it matters: In a graveyard of failed AI wearables, Friend is doubling down on emotional companionship over productivity, betting users will pay premium prices for a digital confidant.
HeyGen just rolled out Video Podcast, a new tool that transforms documents, links, or prompts into fully produced two-host video shows. Instead of stopping at audio, it automatically generates studio scenes, multi-camera switching, and B-roll in minutes. Why it matters: it turns static text into broadcast-ready video content with zero filming required.
Google just updated its Gemini app for macOS with instant voice controls triggered by long-pressing the Fn key. It handles clean dictation by default—stripping filler words on the fly—and can read your screen to summarize highlighted files, rewrite text, or tweak images. It's rolling out globally in English.
A new open-source system called VocalRender generates 48 kHz expressive vocals directly from lyrics and MIDI notes without manual timing alignment. The team released full training code, preprocessing scripts, and two 9.5 GB checkpoints under Apache-2.0. Why it matters: it bypasses tedious phonetic-duration tuning, letting composers render vocal tracks straight from symbolic scores.
xAI just released Grok Voice 2.0, a speech-to-speech model that reasons through queries in parallel while speaking, cutting time-to-first-audio down to 0.7 seconds. It beat GPT-Realtime-2.1 on Artificial Analysis benchmarks and costs $0.08 per minute, making high-speed, human-like voice agents practical for real-world phone calls.
OpenAI just dropped two new transcription models: gpt-transcribe for batch audio and gpt-live-transcribe for real-time speech. Both are designed to tackle messy real-world audio, handling background noise, custom vocabulary, accents, and fast replies. Why it matters: Dependable live transcription is the missing piece for reliable voice agents.
Fish Audio just secured $52M in seed funding and launched S2.1 Pro, a new voice cloning model that requires just five seconds of sample audio. The startup claims it runs twice as fast as Cartesia at one-sixth the cost of ElevenLabs, featuring word-level control over pacing and emotion. Frontier AI firms like HeyGen and LiveKit are already running it in production.
