AssertionError in turn_detector: “end_of_utterance prediction should always return a result”

Description

When running the LiveKit Agents voice bot inside Docker (GCP VM), the turn detector fails as soon as the user speaks after the bot’s first speech. The MultilingualModel().predict_end_of_turn method returns None, triggering an assertion error.

Steps to Reproduce

1. Clone the repository and place the provided Dockerfile, docker-compose.yml, requirements.txt, and code snippet (below) into the project root.
2. Build and start the container:

   docker-compose up --build

3. Wait for the bot to initialize and deliver its first speech.
4. Speak to the bot (first or subsequent utterances).
5. Observe the following error in logs: AssertionError: end_of_utterance prediction should always return a result

Code Snippet (AgentSession Configuration)

from livekit.agents import AgentSession
from livekit.plugins.turn_detector.multilingual import MultilingualModel
# ...
session = AgentSession(
    stt=stt,
    llm=llm,
    tts=tts,
    vad=silero.VAD.load(),
    turn_detection=MultilingualModel(),
    allow_interruptions=True,
    min_interruption_duration=1,
    min_interruption_words=3
)

Relevant Log Excerpt

{"message": "Error in _bounce_eou_task", "level": "ERROR", "name": "livekit.agents", "pid": 25, "job_id": "AJ_mKhPTx564xvC", "timestamp": "2025-05-31T20:24:58.014345+00:00"}
Traceback (most recent call last):
  File ".../livekit/agents/voice/audio_recognition.py", line 321, in _bounce_eou_task
    end_of_turn_probability = await turn_detector.predict_end_of_turn(chat_ctx)
  File ".../livekit/plugins/turn_detector/base.py", line 205, in predict_end_of_turn
    assert result is not None, "end_of_utterance prediction should always returns a result"
AssertionError: end_of_utterance prediction should always returns a result

Environment

• GCP VM: e2-small (2 vCPUs, 2 GB RAM) on Intel Broadwell (x86_64)
• Dockerfile (Python 3.11.6-slim):

ARG PYTHON_VERSION=3.11.6
FROM python:${PYTHON_VERSION}-slim
ENV PYTHONUNBUFFERED=1
ARG UID=10001
RUN adduser --disabled-password --gecos "" --home "/home/appuser" --shell "/sbin/nologin" --uid "${UID}" appuser
RUN apt-get update && apt-get install -y gcc python3-dev && rm -rf /var/lib/apt/lists/*
USER appuser
RUN mkdir -p /home/appuser/.cache && chown -R appuser /home/appuser/.cache
WORKDIR /home/appuser
COPY requirements.txt .
RUN python -m pip install --user --no-cache-dir -r requirements.txt
COPY . .
RUN python main.py download-files
EXPOSE 8081
CMD ["python", "main.py", "start"]

docker-compose.yml

services:
  lk-agent:
    build: .
    ports:
      - "8081:8081"
    env_file:
      - .env
    command: ["python", "main.py", "start"]

requirements.txt

livekit==1.0.8
livekit-agents==1.0.23
livekit-plugins-turn-detector==1.0.23
silero==1.0.23
# …other dependencies as listed

Expected Behavior

The turn detector should always return a valid probability/result when the user speaks, allowing uninterrupted conversation flow

Actual Behavior

MultilingualModel().predict_end_of_turn returns None, causing an AssertionError and halting the voice bot’s processing

Additional Context

• The error does not occur before the bot’s first speech.
• It happens immediately when the user attempts to speak (“end_of_turn” logic never yields a non-None result).
• Disabling turn detection allows the session to continue without crashing.