Ant Ling has open-sourced 5 models in the Ming-3.0 series, covering three speech families: tts, audio, and asr. Ming-3.0-tts includes a 0.5B speech design and paralinguistic cue model and a Tiny 7.9B A1.3B long-speech model; Ming-3.0-audio includes a Flash 126B A6.5B understanding and tool model and a Tiny 9.3B A2.7B efficient dialogue model; Ming-3.0-asr is Tiny 8.6B A2B, aimed at dialects and context.
Ethan Mollick says his new book officially goes on sale tomorrow, and today is the last day to get the preorder bonus. The bonus is an approximately 10-minute AI voice interview that generates personalized prompts, ideas, and reading suggestions based on the book's ideas.
ElevenLabs has released a speech-to-text workflow guide for content creators, covering how to use speech-to-text to generate video captions, podcast show notes, searchable content libraries, and content translations. ElevenLabs Caption Generator can generate transcripts with word-level and character-level timestamps and export SRT and VTT files; the Scribe v2 model transcribes verbatim by default, and you can enable "Clean Transcript" in the Playground or switch to cleaned transcription via the no_verbatim parameter in the ElevenAPI.
ElevenLabs has released a speech-to-text workflow guide for content creators, covering how to use speech-to-text to generate video captions, podcast show notes, searchable content libraries, and content translations. ElevenLabs Caption Generator can generate transcripts with word-level and character-level timestamps and export SRT and VTT files; the Scribe v2 model transcribes verbatim by default, and you can enable "Clean Transcript" in the Playground or switch to cleaned transcription via the
Google Developers Blog published a step-by-step tutorial on how to build a voice agent with ADK and gemini-3.8-live that does not stay silent for long periods. The core approach is to change tool functions from return to yield to make them streaming tools, or to use FunctionTool's response_scheduling to deliver results when idle; in the sales report benchmark flow, the agent's time to first speech dropped from 14.4 seconds to 1.2 seconds. The tutorial also covers the stop_streaming cancellation mechanism, preventing duplicate calls, and the trade-offs among the three scheduling modes WHEN_IDLE, INTERRUPT, and SILENT. It requires google-adk >= 2.10.0 and uses run_live bidirectional sessions.
Google Developers Blog published a step-by-step tutorial on how to build a voice agent with ADK and gemini-3.8-live that does not stay silent for long periods. The core approach is to change tool functions from return to yield to make them streaming tools, or to use FunctionTool's response scheduling to deliver results when idle; in the sales report benchmark flow, the agent's time to first speech dropped from 14.4 seconds to 1.2 seconds. The tutorial also covers the stop streaming cancellation
fal announced that Kotoba's speech API is now live on the platform, supporting real-time speech translation (text and audio output), streaming audio transcription, preset-voice speech generation, and recording transcription with optional word-level timestamps.
AntLingAGI thanks atomic.chat for integrating Ming-3.0 into local AI workflows, making open-source speech models more accessible. According to atomic.chat, Ming-3.0 comprises 5 open-source models covering speaking, listening, and speech transcription; its speech model ranks No. 1 among open-source models on Artificial Analysis, and the smallest version is only 0.5B.
Ant Ling has open-sourced 5 speech models covering TTS, audio understanding, and ASR, with total parameter counts ranging from 0.5B to 126B (6.5B active parameters). vLLM announced that vLLM-Omni provided Ming-3.0 support on the day of release; Ant Ling thanked them for this and said deployment is now more convenient.
Ant Ling posted a thank-you to @AdinaYakup for sharing Ming-3.0 and helping more developers learn about the capabilities of its five open-source speech models. The quoted original post mentioned that Ant Group had just released new speech and audio models.
Deepgram partnered with AWS to port the Nova-3 speech-to-text model to Trainium2, running 8 model replicas on a single trn2.3xlarge instance, achieving an 11%–70% increase in audio processed per dollar compared with the previous solution in offline batch processing scenarios. The team used TorchNeuron to complete the PyTorch port and tuned it through NKILib fused kernels and hand-written NKI kernels, among which the FFN kernel reduced latency by 2.0x, depthwise conv1d sped up by 1.5x, and the conv2d kernel unlocked batch sizes above 32.
Deepgram partnered with AWS to port the Nova-3 speech-to-text model to Trainium2, running 8 model replicas on a single trn2.3xlarge instance, achieving an 11%–70% increase in audio processed per dollar compared with the previous solution in offline batch processing scenarios. The team used TorchNeuron to complete the PyTorch port and tuned it through NKILib fused kernels and hand-written NKI kernels, among which the FFN kernel reduced latency by 2.0x, depthwise conv1d sped up by 1.5x, and the co
Native audio models process or generate acoustic representations directly, preserving information such as timing, intonation, and speaker cues that text transcription would discard. They are fundamentally different from the "transcribe to text first, then synthesize speech" pipeline: the latter changes the causal chain, and the evidence of success, the sources of cost, and the approach to risk control all differ. The article breaks down its workflow into five stages: converting waveforms into audio tokens, modeling speech and non-speech context, inferring semantics and dialogue state, generating audio tokens, and decoding natural speech.
Native audio models process or generate acoustic representations directly, preserving information such as timing, intonation, and speaker cues that text transcription would discard. They are fundamentally different from the "transcribe to text first, then synthesize speech" pipeline: the latter changes the causal chain, and the evidence of success, the sources of cost, and the approach to risk control all differ. The article breaks down its workflow into five stages: converting waveforms into au
ElevenLabs released a guide on how to connect ElevenAgents voice agents to CRM and customer service systems such as Salesforce, HubSpot, and Zendesk, enabling bidirectional data operations such as reading customer records, creating tickets, and updating contacts during calls. The article states that with real-time context access, agents resolve inquiries up to 10 times faster than traditional approaches such as scripted chatbots or fixed phone trees, and can automatically log ticket history and reduce manual data entry through in-call and post-call webhooks.
ElevenLabs released a guide on how to connect ElevenAgents voice agents to CRM and customer service systems such as Salesforce, HubSpot, and Zendesk, enabling bidirectional data operations such as reading customer records, creating tickets, and updating contacts during calls. The article states that with real-time context access, agents resolve inquiries up to 10 times faster than traditional approaches such as scripted chatbots or fixed phone trees, and can automatically log ticket history and
LiveKit has released a guide sharing the experience of its solutions architects and forward-deployed engineers in helping customers move voice agent POCs into production systems, centered on three questions: whether the agent is smart and natural enough, whether production performance can be observed, and whether it can scale reliably and at low cost. The guide offers 7 methods, including measuring first-turn response latency separately, splitting context into the part needed for the greeting and the part loaded in the background, and enabling turn_detection=inference.TurnDetector() and adaptive interruption handling by default.
LiveKit has released a guide sharing the experience of its solutions architects and forward-deployed engineers in helping customers move voice agent POCs into production systems, centered on three questions: whether the agent is smart and natural enough, whether production performance can be observed, and whether it can scale reliably and at low cost. The guide offers 7 methods, including measuring first-turn response latency separately, splitting context into the part needed for the greeting an
