Positioning and Release Date
Seed LiveInterpret 2.0 is an end-to-end real-time simultaneous interpretation model from ByteDance's Seed team. It uses a full-duplex speech understanding and generation framework and is intended for cross-language communication. The official announcement was published on 2025-07-24. (seed.bytedance.com)
Capabilities and Uses
- Supports streaming speech input and generates translations while listening. It can output translated text or speech and supports zero-shot voice cloning based on real-time speech. It is suitable for scenarios such as cross-language meetings, speeches, real-time captions, and voice calls. (seed.bytedance.com)
- The official launch primarily described bidirectional Chinese-English translation. According to the current Volcano Engine service documentation, Doubao Simultaneous Interpretation Model 2.0 supports 48 languages and 9 Chinese dialects; at least one of the source and target languages in a translation direction must be Chinese or English. This scope is described in the current service documentation and should not be taken as the scope supported on the release date. (seed.bytedance.com)
- In the official RealSI Chinese-English evaluation, the average human-rated translation quality score for the speech-to-text task was 74.8/100; the overall score for the speech-to-speech task was 66.3/100. Reported average output latencies were 2.21 seconds and 2.53 seconds, respectively. These results were obtained under specific test conditions and are not guaranteed in all scenarios. (seed.bytedance.com)
Limitations
The official announcement notes that voice cloning stability, speech expressiveness, emotion cloning, and translation accuracy in extremely complex scenarios still have room for improvement. Actual performance and latency are affected by the clarity, fluency, and complexity of the input speech; important content should be reviewed by a human. (seed.bytedance.com)
Pricing
As checked on 2026-11-18, Volcano Engine's domestic site lists the service as “Doubao Simultaneous Interpretation Model,” with pay-as-you-go prices in Chinese yuan (CNY) of 80 yuan/million tokens for input, 80 yuan/million tokens for output text, and 300 yuan/million tokens for output audio. Official documentation states that input audio uses approximately 6.25 tokens per second and output audio approximately 25 tokens per second; actual usage is determined by the bill. These are domestic service prices; the specific applicable conditions and charges are determined by the console. (docs.volcengine.com)
The same domestic pricing document also lists one-year prepaid resource packages: 1 billion tokens for 56,000 yuan, 5 billion tokens for 240,000 yuan, and 10 billion tokens for 400,000 yuan. Input and output text are deducted at a 1:1 ratio, while each output audio token is converted to 3.75 tokens for deduction. The unit price derived from a resource package cannot be treated directly as a uniform price for all types of actual tokens. (docs.volcengine.com)
The simultaneous interpretation pricing page on the BytePlus international site was also searched, but no verifiable pricing information could be obtained in this reading, so no USD amount is listed. (docs.byteplus.com)