Speech-to-speech Response Generation on Synthetic dialogue (test)
4.83Average QualityCascaded
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CascadedBack-end type=GPT-4o back-end, Slow-path LLM=GPT-4o, P90 Latency (ms)=10912026.03 | 4.83 | 5.5 | |
| RelayS2SBack-end type=GPT-4o back-end, Slow-path LLM=GPT-4o, P90 Latency (ms)=812026.03 | 4.78 | 7.4 | |
| CascadedBack-end type=7B back-end, Slow-path LLM=Qwen2.5-7B, P90 Latency (ms)=5132026.03 | 4.38 | 21.3 | |
| RelayS2SBack-end type=7B back-end, Slow-path LLM=Qwen2.5-7B, P90 Latency (ms)=812026.03 | 4.35 | 22.3 | |
| RelayS2SBack-end type=0.5B back-end, Slow-path LLM=Qwen2.5-0.5B, P90 Latency (ms)=812026.03 | 3.36 | 51.4 | |
| CascadedBack-end type=0.5B back-end, Slow-path LLM=Qwen2.5-0.5B, P90 Latency (ms)=4202026.03 | 3.34 | 51.8 | |
| S2S onlyBack-end type=0.5B back-end, Slow-path LLM=—, P90 Latency (ms)=712026.03 | 3.04 | 59.3 |