Skip to content

Seed-ASR

An LLM-based speech recognition model

Visit official site

What Seed-ASR is

The model is described as an LLM-based speech recognition model developed by ByteDance, feeding continuous speech representations and context information into the LLM. It is presented as improving results across domains, accents, dialects and languages through large-scale training and context awareness.

What you can do with it

  • Recognize speech by combining continuous speech representations with context
  • Use dialogue history, agent names and agent descriptions to raise accuracy
  • Recognize Chinese, English and dialects such as Wu, Cantonese and Sichuanese
  • Reuse user subtitle edits as hints to avoid the same error in later videos

Application scenarios

  • An enterprise transcribes meeting recordings in real time
  • A content creator turns video or podcast speech into text
  • An education institution transcribes classroom recordings for review

Getting started

  • Visit the official site or download the related app
  • Register, log in and choose a suitable service plan
  • Upload a speech file or run real-time recognition directly
  • Set recognition parameters such as language and dialect

Pros & cons

✓ What we like

  • User edits to subtitles act as hints to avoid repeated errors
  • Accurate recognition across business, education and entertainment domains
  • Chinese, English and several Chinese dialects are supported

! What to watch out for

  • Recognition is built around a large language model and context input
  • Language and dialect choices are set as recognition parameters

FAQ

Who developed Seed-ASR?

The material says it was developed by ByteDance and is based on a large language model.

What result is reported on public test sets?

The material states a 10%-40% word error rate reduction on Chinese and English public test sets.

Which dialects are named?

Wu, Cantonese and Sichuanese are listed among the Chinese dialects.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review