Deepgram speech, Pipecat voice agents, fal video via MCP, and DatI database gateways make every modality callable

Speech that handles the real world

Deepgram's update expands its speech models in the directions real conversations actually go. The company says the new Nova-3 multilingual model supports 55 languages within a single model, supports code-switching for speech that switches languages mid-conversation, and lets users bias language detection using the optional language_hint parameter. Nova-3 multilingual responses now include a languages_hinted field. The details matter: a multilingual model that cannot handle a speaker switching languages mid-sentence is a demo, not a product, and a hint parameter is how a developer tells the model what the conversation is likely to be. m:…3309

On the synthesis side, the company says Flux TTS has added 41 new languages and 230 new voices, with text normalization available for 9 languages, while 32 Flux TTS languages are currently in Early Access without text normalization. New Flux TTS voices are available on the company's main, European, and Indian API endpoints, but not yet on the Australia endpoint. The Early Access caveat is worth noting: 32 languages ship without text normalization, which means numbers, dates, and abbreviations in those languages may not be spoken correctly yet. A developer shipping in those languages should test that path before promising it. m:…3309

Voice agents get new services and a deprecation warning

Pipecat's v1.14.0 release shows a voice-agent framework absorbing the agent era. The release adds PerplexityAgentHttpLLMService for Perplexity's Agent API and NvidiaOmniLLMService for NVIDIA Nemotron Omni models, and adds text_transforms to LLMTextProcessor for rewriting output aggregations. SpeechmaticsSTTService now supports 8 kHz audio, and TwilioFrameSerializer now supports authentication with a Twilio API key. NvidiaLLMService now defaults to nvidia/nemotron-3.5-super-vl-120b-a12b. The release is, in effect, a tour of the current voice stack: search-enabled agents, multimodal models, and telephony plumbing, all in one framework. m:…8225

The release also carries the maintenance news developers actually need. PerplexityLLMService is deprecated in favor of PerplexityAgentHttpLLMService, WebsocketService.send_with_retry is deprecated and will be removed in 2.0.0, and the minimum version for aiohttp is raised to 3.14.4 to address security vulnerabilities. Vertex AI will retire gemini-3.6-flash on 2026-11-19, according to the release notes. Deprecations are the unglamorous part of a framework release, and they are the part that breaks production systems if ignored. A team running Pipecat should treat this release as a migration checklist, not just a feature list. m:…8225

Video generation and databases join the same pattern

The fal tutorial shows video generation becoming a chat capability. Connecting the fal MCP server to Claude allows Claude Opus 5.5 to write and run Seedance 2.5 prompts directly from a chat, according to the tutorial, and Claude Code is the preferred route when product photos are saved as local files on a computer. Seedance 2.5 drafts render at 480p and can be completed at 1080p within seven days, with 480p drafts priced at $0.2149 per second and 1080p completions billed at approximately $1.164 per second. The fal MCP server adds no additional cost on top of standard model run rates, according to the tutorial, and Seedance 2.5 requests return one MP4 file of up to 30 seconds with an integrated soundtrack. Reference-to-video accepts up to 30 product photos, and Claude Code can upload local files up to 90 MB to fal using a signed upload URL. m:…8742

DatI completes the interface layer on the data side, as described in the agent-stack story: it is a lightweight semantic gateway connecting AI agents to enterprise databases, publishing databases as MCP services using Streamable HTTP, with schema-level permission control and automatic registration of metadata and semantic indices for newly created tables. Agents can create intermediate or analysis tables within authorized schemas. The pattern across all four pieces is the same: give the agent a callable interface for each modality, with the permissions and pricing visible. Speech, video, voice frameworks, and databases are no longer separate product categories. They are the agent's senses and hands, and the developers who wire them together with the right permissions are the ones shipping working agents. m:…8370