A compact in-browser speech model that streams audio fast and reports word timings for live highlighting.
“Browser TTS with Per-Word Highlight Timings on WebGPU” is an advanced Software & DevOps AI agent workflow from Software. It connects Claude and Other or custom in 4 steps. Setup takes about Not stated and running costs are about Not stated.
Running text-to-speech locally in a browser usually means either a large download or no word-level timing data, which makes synced highlighting hard. The official runtime for a small speech model was heavier than needed and did not expose per-word timings, so a second runtime was built to keep generation on the GPU and read timings from the model itself while it speaks.
Convert every large matrix in the speech model, including the audio decoder, to int8 so the browser download shrinks well below the official package size.
Implement the full generation loop as GPU kernels so audio synthesis never round-trips to the CPU during inference.
Read timing information from one of the model's attention heads as it generates speech, then align each word to its audio span for highlighting.
Start playback as soon as the first audio chunk is ready, and compare throughput and timing accuracy against the reference runtime on desktop and mobile.
You are helping me port a small text-to-speech model to run fully in the browser on WebGPU. Outline a plan for: storing all large weight matrices as int8, including the audio decoder; writing generation kernels that keep the loop on the GPU; and exposing per-word timings read from an attention head during synthesis. Assume the model is [MODEL NAME] and the target is [BROWSER TARGET]. List the kernel stages in order, note where quantization happens, and describe how to stream the first audio chunk quickly. Do not include code yet; give me the architecture and the risks to check first.
I have a browser-based speech runtime on WebGPU and want reliable word-level timings for live highlighting. Review this approach: read timing signals from one attention head while audio is generated, map them to word boundaries, and emit them alongside streamed audio chunks. Point out failure modes such as drift, silence handling, and punctuation, and suggest validation checks against [REFERENCE AUDIO SET]. Keep the response focused on correctness and latency, and propose a small test matrix across [DEVICE LIST].
A compact in-browser speech model that streams audio fast and reports word timings for live highlighting.
You need Claude and Other or custom.
Plan for about Not stated. It is rated advanced. The guide walks through 4 steps, starting with “Store weights as int8”.
Expect about Not stated in running costs (Free range).
Running text-to-speech locally in a browser usually means either a large download or no word-level timing data, which makes synced highlighting hard. The official runtime for a small speech model was heavier than needed and did not expose per-word timings, so a second runtime was built to keep generation on the GPU and read timings from the model itself while it speaks.
Score every demo request against your ICP, update HubSpot and ping the right rep in Slack within a minute of the form submit.
Every new Zendesk ticket gets a category, priority and sentiment tag plus a draft reply in an internal note before an agent opens it.
Copilot and Claude review each PR for bugs, missing tests and risky changes before a human looks, so reviewers can focus on design.
Built this workflow? Show it is featured on AgentUseHub with a live upvote count that links back here. About badges
