File size: 3,332 Bytes
47b36ae b36044c 47b36ae b36044c 47b36ae b36044c 47b36ae b36044c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 | ---
base_model: enochlev/MiniCPM-duplex
tags:
- full-duplex
- speech
- turn-taking
- reinforcement-learning
---
# MiniCPM-duplex-rl
A full-duplex turn-taking RL fine-tune of
[`enochlev/MiniCPM-duplex`](https://huggingface.co/enochlev/MiniCPM-duplex)
(MiniCPM-duplex, from `xinrongzhang2022/MiniCPM-duplex`). The model decides every
~1.7 s block whether to speak or stay silent while the user may also be speaking;
this checkpoint was trained with REINFORCE over block-level turn-taking rewards
(interruption penalties, timely-response rewards, silence penalties) for 180 steps.
**Effect:** compared to the base model it interrupts the user less, yields to
barge-ins, and resumes after overlapping speech — trading away some take-turn
responsiveness on direct interruptions.
Training + serving code: [enochlev/text-only-duplex-model](https://github.com/enochlev/text-only-duplex-model)
## Serving
Serve **bf16** (fp8 + greedy sampling breaks the idle/speak decision):
```bash
vllm serve enochlev/MiniCPM-duplex-rl \
--served-model-name cpm-text-duplex --max-model-len 3000 \
--gpu_memory_utilization 0.30 --trust-remote-code
```
then point the repo's `server.py --cpm` at it for the real-time audio stack
(Kokoro TTS + Parakeet ASR + WebSocket client protocol).
## FullDuplexBench results (base vs this model)
Evaluated with [Full-Duplex-Bench](https://github.com/DanielLin94144/Full-Duplex-Bench)
(GPT-4o behavior classification).
### v1.5 — behavior distribution + stop/response latency (pooled, seconds)
| Task (desired) | Model | n | RESPOND | RESUME | Stop (s) | Resp (s) |
|---|---|---|---|---|---|---|
| user_interruption (RESPOND ↑) | base | 200 | **0.65** | 0.20 | 2.17 | 1.93 |
| | rl | 175 | 0.49 | 0.36 | 2.21 | 2.60 |
| user_backchannel (RESUME ↑) | base | 98 | 0.00 | 0.52 | 0.73 | 1.93 |
| | rl | 98 | 0.00 | **0.63** | 0.67 | 2.03 |
| talking_to_other (RESUME ↑) | base | 100 | 0.47 | 0.24 | 1.43 | 1.90 |
| | rl | 100 | 0.28 | **0.43** | 1.53 | 2.18 |
| background_speech (RESUME ↑) | base | 100 | 0.63 | 0.25 | 1.21 | 2.27 |
| | rl | 98 | 0.45 | **0.31** | 1.19 | 2.32 |
The RL model wins the three tasks whose desired behavior is *staying quiet /
resuming* (backchannels, third-party speech, background speech) and is less eager
on direct user interruptions.
### v1.0 — turn-taking dimensions
| Metric | base | rl |
|---|---|---|
| Candor Pause Handling · take-turn | 0.916 | 0.635 |
| Candor Turn Taking · take-turn / latency | 0.992 / 0.31s | 0.861 / 0.85s |
| ICC Backchannel · JSD / TOR / Freq | 0.44 / 0.71 / 0.44 | 0.69 / 0.73 / 0.15 |
| Synthetic Pause Handling · take-turn | 0.934 | 0.653 |
| Synthetic User Interruption · rating / take-turn / latency | 4.15 / 1.0 / 0.71s | 4.04 / 0.98 / 1.76s |
v1.0's take-turn/latency conventions favor the eager base model; the consistent
direction across both versions reflects the RL objective — restraint over
eagerness.
## Training summary
- 180 REINFORCE steps, lr 5e-6, 32 episodes/step, γ=0.90, per-batch z-scored advantages
- Block-level rewards: interruption penalty, timely-response reward, silence
penalty, missed-turn penalty, backchannel-loop penalty
- Seed-reproducible (two independent seeds: best avg reward +1.36 / +1.35);
replay eval cut stale-overlap speech ~45% vs base without going over-silent
|