Whisper

Beyond Transcription: Building Voice AI That Understands Conversations — Hervé Bredin, pyannoteAI

Jun 5, 2026 · 25:20

Hervé Bredin, chief science officer at pyannoteAI, argues that speaker diarization benchmarks are misleading because they use headset audio while users rely on table microphones—Nvidia Parakeet reports 11.4% word error rate on AMI headset data but gets 26% on the same dataset's table mic. The episode covers how speaker diarization (who speaks when) is harder than it looks, especially when combining with transcription: overlapping speech, timestamp disagreements, and words falling between speaker boundaries create errors. Bredin demonstrates pyannoteAI's Precision 2 model achieving 3% diarization error rate (DER) against a 5% baseline on a two-speaker phone call. State of the art today: 2% DER on clean telephone calls but 41% in a noisy restaurant, showing the problem is far from solved. The reconciliation between diarization and STT is handled by a proprietary orchestration that works with any STT model.

Engineering voice agents: Latency, quality, and scale — Rishabh Bhargava, Together AI

May 31, 2026 · 24:35

Rishabh Bhargava from Together AI outlines engineering voice agents, explaining that pipeline architecture with colocated models can achieve sub-500ms response times critical for user retention. Speech-to-text targets P90 under 100ms and 6% word error rate, while the LLM must stay within 200-300ms time-to-first-token using 8-30B parameter models—larger models blow the budget, smaller ones break tool calling. Network latency from distant data centers adds 75ms (30% overhead) versus 5ms when colocated in the same building. Pure speech-to-speech models are emerging but still struggle with instruction following and tool calling. The thinker-talker pattern uses a small LLM for fast conversational flow and issues a single tool call to a larger model for complex requests.

Building voice agents with OpenAI — Dominik Kundel, OpenAI

Jun 29, 2025 · 1:25:35

Dominik Kundel from OpenAI presents the new OpenAI Agents SDK for TypeScript and demonstrates how to build voice agents, arguing that speech-to-speech architectures (using GPT-4 real-time) offer lower latency and more natural interactions than chained approaches that transcribe audio to text. He covers two architectures: chained (speech-to-text, text agent, text-to-speech) versus speech-to-speech (native audio understanding), and recommends starting with a small, clear goal, building evals and guardrails early, and using generative prompts to control tone and personality via openai.fm. In a live coding session, he builds a real-time agent with tools (e.g., getWeather) and handoffs, demonstrates delegation to a smarter model (O4 mini) for complex tasks like refunds, and shows built-in interruption handling, human-in-the-loop approval, tracing for debugging, and output guardrails. The SDK supports WebRTC and WebSocket, handles turn detection automatically, and provides session management with a 30-minute timeout that can be extended by injecting previous context.

Realtime Conversational Video with Pipecat and Tavus — Chad Bailey and Brian Johnson, Daily & Tavus

Jun 27, 2025 · 18:46

Chad Bailey of Daily and Brian Johnson of Tavus explain how to build real-time conversational video bots using the Pipecat open-source framework and Tavus's avatar platform. They argue that beyond models, an orchestration layer is essential for handling input, processing, and output with low latency. Bailey details Pipecat's pipeline of frames, processors, and pipelines that manage audio/video frames, speech-to-text, LLM inference, and text-to-speech in a modular way. Johnson describes Tavus's 600-millisecond response time and proprietary models—Sparrow Zero, Raven Zero—plus future turn detection, response timing, and multimodal perception models being integrated into Pipecat. They emphasize that Pipecat solves real-world production challenges like observability, barge-in handling, and parallel pipelines for tasks like voicemail detection. Johnson admits Tavus initially built its own orchestration but now plans to adopt Pipecat internally, as its customers are already using it.

Text-to-Speech Data Preparation and Fine-tuning Workshop - Ronan McGovern

Jun 3, 2025 · 34:00

Ronan McGovern walks through fine-tuning Sesame's CSM-1B text-to-speech model on a specific voice, using a YouTube video as the data source. He explains token-based TTS models, including how audio is represented via codebooks and how CSM-1B uses a main transformer for zeroth tokens and a secondary transformer for 31 hierarchical tokens. The workshop covers data preparation: downloading audio with yt-dlp, transcribing with Whisper Turbo, manually correcting the transcript, and splitting audio into 30-second chunks (41 clips from a 30-minute video). Fine-tuning uses Unsloth with LoRA adapters (rank 32, alpha 16) on linear layers, training for one epoch with a batch size of 2 and virtual size of 8, reducing loss from ~6.34 to ~3.72. Evaluation compares zero-shot inference (random speaker), voice cloning (closer but imperfect), and fine-tuned plus cloning (best result, producing an Irish-accented voice with natural errors). McGovern recommends 50+ 30-second clips for noticeable effect and notes that combining fine-tuning with voice cloning yields good performance even with limited data.

Voice Agents: the good, the bad, and the ugly

Feb 22, 2025 · 18:48

Eddie Siegel, CTO at Fractional AI, details the real-world challenges of building an AI voice agent for consulting-style interviews, moving from proof of concept to production. The agent, designed to conduct hundreds of research interviews simultaneously, struggled with hallucinations, latency, and Whisper transcription errors like 'creeping dippity dippity dippity dippity'. To constrain the LLM's tendency to chitchat and rabbit-hole, Siegel's team introduced a 'drift detector' agent and a 'next question' agent, both running as separate text-based side threads. They also added tool use for question navigation, explicit goals and priorities for follow-ups, and a 'transcript cleaner' to hide garbled outputs from the user interface. To move beyond vibes-driven iteration, they built an automated eval suite with metrics like clarity and completeness, using synthetic conversations with personas such as a 'snarky teenager in charge of a Fortune 500 company' to simulate diverse interviewees. The talk offers a no-nonsense playbook for getting voice agents into production, focusing on out-of-band checks, tool use, and evals even without objective ground truth.

Multi model multimodal and multi agent innovations in Azure AI: Cedric Vidal

Feb 6, 2025 · 28:56

Cedric Vidal, Principal AI Advocate at Microsoft, demonstrates Azure AI's multi-model, multimodal, and multi-agent capabilities in a session packed with live demos. He shows how GPT-4 Omni mixes text and vision to read handwritten French menus and translate them, and diagnoses infrastructure damage from photos for energy and insurance industries. A new video translation service lets him speak German, Spanish, Italian, and Japanese in his own voice while preserving tone, such as whispering or yelling. The Azure AI model catalog now offers 1,600 models, including serverless deployment options, and Phi-3 Vision, a small 3.8B parameter model running locally in a browser via WebGPU. He also demonstrates code interpreter analyzing a GPX file from a kitesurfing session, plotting a map with turn markers, and GitHub Workspaces preview generating a Java GUI from Python code.

Productionizing GenAI Models – Lessons from the world's best AI teams: Lukas Biewald

Oct 23, 2024 · 22:36

Lukas Biewald, founder of Weights and Biases, shares lessons from productionizing GenAI models, emphasizing that while AI is easy to demo, it is hard to productionize. He illustrates this with a personal project building a custom Alexa-like device using LLaMA and Whisper, where accuracy improved from 0% to 98% through prompt engineering, switching to Mistral, and fine-tuning with QLoRA. Biewald argues that tracking experiments (including failures) is critical for reproducibility and collaboration, and that a robust evaluation framework—beyond 'testing by vibes'—is essential for iterating and shipping v2. He notes that 70% of the audience had LLM apps in production, yet many lack solid evaluations, and recommends starting with lightweight prototypes and incorporating end-user feedback.

From model weights to API endpoint with TensorRT LLM: Philip Kiely and Pankaj Gupta

Sep 13, 2024 · 1:40:01

Philip Kiely and Pankaj Gupta of Baseten lead a workshop on TensorRT-LLM, NVIDIA's high-performance inference framework for LLMs, arguing its use delivers best-in-class throughput and latency on NVIDIA GPUs. They explain that TensorRT-LLM optimizes computational graphs via plugin kernels and in-flight batching, achieving 216 tokens per second and 180ms time to first token on Mistral 7B. The workshop demonstrates building an engine for TinyLlama 1.1B, including FP8 quantization that reduced engine size from 2 GB to 1.2 GB with minimal quality loss. They benchmark a deployed model, showing 7,000 total tokens per second at batch size 64 on an A10G. The presenters compare TensorRT-LLM favorably to VLLM for high-throughput production use and introduce Truss, Baseten's open-source packaging tool, alongside their managed platform for automatic scaling and fast cold starts.

From Text to Vision to Voice Exploring Multimodality with Open AI: Romain Huet

Jul 10, 2024 · 23:39

OpenAI's Romain Huet demonstrates GPT-4o's real-time audio, vision, and screen-sharing capabilities at the AI Engineer World's Fair, showing how the omni-model achieves human-like conversational latency, understands emotion, and can be interrupted naturally. He walks through live demos of ChatGPT Desktop recognizing drawings, reading book pages, and pair-coding a responsive travel app using Tailwind CSS. Huet also previews Sora's text-to-video generation and OpenAI's unreleased Voice Engine, which can clone a voice from seconds of audio and translate it into multiple languages while preserving the original speaker's tone. He outlines OpenAI's four focus areas: advancing textual intelligence, faster/cheaper models (GPT-4o is twice as fast as GPT-4 Turbo at half the price), model customization via fine-tuning, and enabling multimodal agents. The talk emphasizes that today's models are the dumbest they'll ever be and urges developers to build for a future where AI works across text, voice, and video.

[Workshop] AI Engineering 101

Nov 6, 2023 · 3:02:24

In this hands-on workshop, AI engineer Noah Hein teaches the basics of AI engineering by building five projects with OpenAI's GPT-3, GPT-4, Dall-E, Whisper, and Telegram. Participants create a Telegram bot that uses GPT-3 for chat, implements retrieval-augmented generation (RAG) with embeddings and cosine similarity on MDN docs, generates code with GPT-4 using few-shot prompting, creates images via Dall-E 2, and transcribes voice with Whisper. The session explains core concepts like tokens, chunking, context windows, temperature, and top-p, emphasizing that embeddings and prompt engineering are key to performance. Hein demonstrates that these tools are cheap and accessible—the entire workshop costs about $0.05 in API fees—and shows how AI engineers can integrate multiple models into a single application.

See, Hear, Speak, Draw: Logan Kilpatrick & Simón Fishman

Oct 24, 2023 · 18:43

OpenAI's Simón Fishman and Logan Kilpatrick argue that 2024 will be the year of multimodal models, showcasing how text can currently bridge vision, audio, and image generation. They demo a DALL-E vision loop where GPT-4 with vision describes an image, DALL-E 3 generates a synthetic version, then the models compare and iterate to improve accuracy—all in roughly 50 lines of code. Another demo extracts frames from a video, uses GPT-4 with vision to describe them, combines with Whisper's transcript, and generates a blog post with matching DALL-E images, making video content more accessible. They emphasize that while today's multimodal capabilities are separate, future unified models will directly reason across modalities, eliminating the need for manual integration.