Search “Speech”

Ollama Launches Speech Model Phonon, Supporting Text-to-Speech

Ollama can now convert text to speech. Its first speech model is the Phonon series of small text-to-speech models from Gradium. It is currently available on macOS, with Windows and Linux support coming soon. On a MacBook Pro equipped with M5 Max, Ollama, with the help of MLX acceleration, is 35% faster than Gradium's own runtime, and its word error rate is on par with the Hugging Face Transformers reference implementation. Phonon is compatible with OpenAI's /v1/audio/speech interface, so existing OpenAI SDK code only needs to change the base URL; phonon is released under Apache 2.0, and the 40M model uses the CC BY-NC 4.0 license and is for non-commercial use only.

Block releases speech model Dessa, trained only on human conversation and able to run locally

Block releases speech model Dessa, saying it is trained only on human conversation, without using text or pretrained models. Dessa uses a full-duplex approach, allowing it to keep listening while speaking and to interject, laugh, or respond; Block says it can run locally on MacBook Pro models with M4 Max and above chips. Dessa is currently a research preview, with no release date or pricing, and early access is open only to a small number of AI safety researchers.

Scammers used AI to clone a daughter's voice and fake a kidnapping; voice experts recommend that families set up a safety code word

Fortune reported that Deborah Del Mastro, a mother in Martinez, California, received a call in May from someone posing as a Mexican drug cartel; the caller played an AI-cloned recording of her daughter crying and screaming. Following instructions, she wired $5,400 to Mexico from multiple locations within five hours, only to later discover that her daughter had been at work the whole time. McAfee tests show that a mere three-second voice sample is enough to generate a clone that matches the original voice by 85%; a Gallup and Stop Scams Alliance survey estimates that in 2025 about 15 million American adults were scammed, with losses of about $68 billion, 12% of which involved AI or deepfakes. Kimberly Hyun, an assistant professor at the University of Cincinnati, recommends that family members agree on a code word known only to each other, and that when receiving a call from an unknown number claiming a family member is in trouble, they should hang up immediately and call that person back directly to verify.

Block releases Dessa, a full-duplex model trained purely on speech, which can run locally in real time on a MacBook Pro M4 Max

Block, founded by Twitter co-founder Jack, has released Dessa, a speech-to-speech model trained from scratch purely on human speech, with no text and no pretrained base, supporting full duplex so it can listen and speak at the same time. The model claims inference faster than real time and can run locally on a MacBook Pro M4 Max and above, and the entire model is said to have been trained on a single NVIDIA GB200 NVL72. The team says its goal is to let the model understand tone and emotion, avoiding the safety risk of continuing to respond without perceiving the other person's distress.

Block launches full-duplex speech model Dessa, supports local operation

Block has launched the full-duplex speech model Dessa, which can listen to the user's speech while speaking, and can simulate emotional expressions such as laughter, pauses, and choking up. The company says it is trained entirely on human speech data, using only a single rack of NVIDIA GPUs for training, and can run locally on MacBooks with M4 Max and higher configurations. At this stage Dessa is not open for public trial; Block is sending testing invitations to security researchers, focusing on risk scenarios in which speech agents deceive humans.

Block releases full-duplex voice model Dessa, which can run locally on consumer-grade GPUs

Block's research lab has released the voice model Dessa, which uses a full-duplex architecture, can listen and speak at the same time, and can run locally on a MacBook M4 Max or a more powerful consumer-grade computer. The model was trained entirely on human speech and a single NVIDIA GPU, and the company says expressions such as emotion, pauses, and laughter are built into the model rather than driven by a text pipeline. Dessa is not currently publicly available, and Block is recruiting security researchers to test the risk that it could be used to deceive conversation partners.

Amazon Audible improves audiobook speech synthesis with controllable TTS

The Amazon Audible technical team adopted controllable TTS, using natural language descriptions to control dialogue performance, enabling audiobooks to distinguish narration from dialogue and identify speaking characters. Its BookLLM achieved 97% and 94% accuracy in identifying speaking characters on 25 English and Spanish books, respectively, and the distinctiveness of character dialogue improved by 3.7 points in MUSHRA tests. In English and Spanish tests, eight of ten voices showed statistically significant improvements in perceived quality.

Google Maps CarPlay version adds Gemini voice conversation feature

Google has added Gemini voice mode to the Google Maps CarPlay version for iPhone users, callable directly starting from Maps version 26.43.1, for querying waypoints, destinations, traffic conditions, and attractions, among others. This feature was previously available on Android Auto; this rollout only states "the United States and other regions of the world" without giving a timeline. The feature is tied to a Google account and can use personal data such as conversation history to adjust answers.

Streaming low-latency speech for AI agents with Kokoro TTS on Amazon SageMaker AI

AWS demonstrated a solution for hosting the Kokoro-82M text-to-speech model on Amazon SageMaker AI that lets AI agents stream-synthesize speech immediately as each sentence is generated, so playback does not have to wait for the full response to finish generating. Kokoro-82M has only 82 million parameters and can run on a single ml.g5.xlarge GPU instance; the container provides two routes, POST /invocations and WS /invocations-bidirectional-stream. The article also explains how binary WebSocket frames are handled in bidirectional streaming, and precautions for keeping the session from closing.

Amazon Polly generative engine major upgrade: adds 14 new voices and supports Hindi and Finnish for the first time

The Amazon Polly generative engine has completed an upgrade, adding 14 new voices across 9 locales and, for the first time, officially offering Hindi and Finnish in the generative engine. After the upgrade, it offers a total of 57 generative voices, covering 11 languages and 25 locales, and has expanded to AWS GovCloud (US-West), for a total of 11 AWS Regions. Existing voice names and API calls remain unchanged; this is an in-place upgrade with no price change and no code changes required.

Sierra launches the 𝜏-Entity voice agent benchmark: GPT Live 1 passes only 34.0% of tasks under all conditions

Sierra releases the 𝜏-Entity benchmark, using 200 tasks and ten entity categories (names, phone numbers, addresses, emails, etc.) to test whether voice agents can accurately capture, recognize, and correct detail errors. On the six-language leaderboard, GPT Live 1 leads, with 60.3% accuracy on regular calls, but only 34.0% of tasks pass under all three conditions—regular, heavy noise, and strong voice; Gemini 3.8 Live and Grok Voice Think Fast 2.0 score 19.5% and 24.0%, respectively. The paper also finds that mandating a "ask—spell out/repeat—confirm correction—submit" workflow can raise the all-condition success rate by 14–31 percentage points, at the cost of 21–28 extra seconds per call.

CarPlay version of Google Maps integrates Gemini, supporting voice navigation and traffic reporting

Google has begun rolling out Gemini to users in Google Maps on Apple CarPlay. A Gemini star icon will appear on the right side of the navigation interface; tapping it allows users to issue multi-step, open-ended commands by voice. Users can ask about scenic spots, the current crowd levels at a location and quieter times, directly request to avoid toll roads when giving an address, and report traffic conditions such as congestion and accidents while driving. The feature launches first in the United States and will roll out to other regions worldwide in the coming weeks.

OKA launches Mori, a mushroom-shaped robot: not a voice assistant, just companionship

OKA has launched Mori, a $649 mushroom-shaped home robot lamp with a built-in camera and microphone that can follow users around, play nature sounds, and guide breathing relaxation, but it does not support voice commands to control smart home devices or check the weather. Mori is powered by multiple local large language models and can judge when to interact and when to stay quiet; images and audio are discarded immediately after local processing, and facial data is stored encrypted and never uploaded. The founder's edition is limited to a few hundred units, while the standard edition costs $799 and does not include subscription features.

MiniMax M3.1 Flash Teams Up with SGLang Omni to Build an Interruptible Voice Agent

MiniMax M3.1 Flash combined with SGLang Omni is used to build voice agents that can be interrupted, corrected, and redirected in real time. M3.1 Flash provides fast, concise response capabilities, while SGLang Omni adds streaming output, dialogue state, and interruption handling. A developer replaced the local LLM in the original Qwen3-ASR → Qwen3 LLM → Qwen3-TTS pipeline with the M3.1 Flash API, leaving the rest unchanged; in testing, responses were fast and answers were concise and direct.

HeyGen Voice Turbo Tops the Artificial Analysis Controlled Voice TTS Arena

HeyGen Voice Turbo ranks first in the Artificial Analysis Controlled Voice TTS Arena with an Elo of 1,202, surpassing the previous-generation HeyGen Voice at 1,196 and Qwen-Audio-3.1-TTS-Plus at 1,185. The arena evaluates using the same set of 8 cloned voices (4 American, 4 British); HeyGen Voice Turbo ranks first among 46 models and ranks first in the Assistants and Entertainment categories. Its price is $4/1M characters, the lowest on the leaderboard, 7.5 times cheaper than the previous-generation HeyGen Voice at $30/1M characters.

Sierra launches 𝜏-i18n multilingual voice agent evaluation benchmark

Sierra releases 𝜏-i18n, extending 𝜏-Voice to five languages: Spanish, Brazilian Portuguese, Hindi, Korean, and Mandarin, retaining the same airline, retail, and telecom tasks, policies, tools, and success criteria to enable cross-language comparison. Across 2,700 main calls, task completion rates for Spanish, Hindi, and Portuguese lag English by an average of 2.4, 3.8, and 6.4 percentage points, while Korean and Mandarin lag by 22.6 and 9.1 percentage points respectively; Korean is the weakest language among the three agents. In addition to task completion, 𝜏-i18n separately reports interaction metrics (response timing, selectivity, tool-call efficiency) and generation quality (naturalness, speech fidelity), and has open-sourced the codebase and the language pack framework LanguageFactory.

How to integrate the ElevenLabs text-to-speech API into an app

ElevenLabs has published a tutorial on how to integrate the text-to-speech API into an app using Python and TypeScript: create an API key, install the ElevenLabs SDK, send text to the TTS endpoint and specify a voice and model, then play the returned audio. The example uses the eleven_v4 model, and the low-latency Eleven Flash v2.5 is also available; the Voice Library offers more than 17,500 voices, and output defaults to mp3 but can be switched to ulaw or pcm.

How Acuity Health Transformed Patient Services with LiveKit Voice Agents

Acuity Health built a custom voice agent on LiveKit: 80% of incoming calls are fully handled by the AI agent, most appointments are completed within 1 minute, and 81% of appointment calls end with "thank you." Its agent typically runs on a Gemma 3 27B model optimized on LiveKit Inference, paired with turn-detection models tuned for different patient populations, supports voice and text, and connects to call center tools, task management software, analytics providers, and EMR systems. The team had previously been constrained by low-code platforms; after switching to LiveKit, they built a working prototype within a week, reached a development environment within a month, and use LiveKit Simulations to stress-test scenarios before every change goes live.

English translation is being prepared.

Ethan Mollick: Voice interfaces may put an end to the open office

Ethan Mollick believes that voice interfaces in AI applications may finally put an end to the "monster" that is the open office—or may simply make it worse. Citing his own research, he points out that multiple experiments show that background noise in open offices reduces the ability to do analytical work, increases software bugs, and weakens the ability to find bugs.