# kyutai-labs/delayed-streams-modeling

Kyutai's Speech-To-Text and Text-To-Speech models based on the Delayed Streams Modeling framework.

Repository: https://github.com/kyutai-labs/delayed-streams-modeling
Canonical: https://ross.abutalabs.com/products/delayed-streams-modeling
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-01-26T09:28:20+00:00

## Health v2 (maintenance only)
Score: 47/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 64, release rhythm 35, longevity 31
- inputs: {"age_days": 443, "days_push": 219, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3017, forks 313 (observed 2026-08-28T04:07:38.390205+00:00)

## What it is
Kyutai's repository of Speech-To-Text and Text-To-Speech models built on the Delayed Streams Modeling framework, with implementations in PyTorch, Rust, and MLX. The models support streaming, real-time inference with word-level timestamps and semantic voice activity detection, suitable for voice agents.

## Use cases
- transcribe live audio streams in real time
- build a voice agent with speech-to-text and text-to-speech
- run speech recognition on-device on a Mac or iPhone
- serve many concurrent transcription streams over websockets
- convert text to natural-sounding speech
- get word-level timestamps from spoken audio

## When to choose
- you need low-latency streaming STT or TTS rather than batch transcription
- you want to build interactive voice AI applications or agents
- you need efficient batching of many simultaneous audio streams on GPU
- you want on-device inference on Apple silicon via MLX

## When to avoid
- you only need offline batch transcription of recorded files with no latency constraints
- you need STT for languages beyond English and French
- you cannot run GPU inference and need a lightweight CPU-only solution
- you need a fully managed cloud speech API rather than self-hosted models

## Facets
- artifact type: library
- maturity: active
- function: speech-recognition, tts, machine-learning, llm-inference
- domain: speech-processing, artificial-intelligence
- platform: python, rust, cli
- tags: streaming-inference, voice-agents, word-level-timestamps, voice-activity-detection, mlx, pytorch, websocket-server, real-time-transcription, audio, macos, ios, gpu, web-server

## Member repositories
- kyutai-labs/delayed-streams-modeling (main) score 47

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:38.390205+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T07:30:03.203512+00:00, confidence not recorded.
  - readme: https://github.com/kyutai-labs/delayed-streams-modeling (fetched 2026-08-28T04:07:38.390205+00:00, sha 48ceee5d5ad7)
- Data as of 2026-08-30T08:39:29.467469+00:00.