Add a Qwen3-ASR plugin for self-hosted vLLM

livekit-agents/pyproject.toml

@@ -111,6 +111,7 @@ openai = ["livekit-plugins-openai>=1.8.5"]
 palabra = ["livekit-plugins-palabra>=1.8.5"]
 perplexity = ["livekit-plugins-perplexity>=1.8.5"]
 protoface = ["livekit-plugins-protoface>=1.8.5"]
+qwen-asr = ["livekit-plugins-qwen-asr>=1.8.5"]
 resemble = ["livekit-plugins-resemble>=1.8.5"]
 respeecher = ["livekit-plugins-respeecher>=1.8.5"]
 rime = ["livekit-plugins-rime>=1.8.5"]

livekit-plugins/livekit-plugins-qwen-asr/README.md

@@ -0,0 +1,58 @@
+# Qwen3-ASR plugin for LiveKit Agents
+
+Speech-to-text for a self-hosted [Qwen3-ASR](https://github.com/QwenLM/Qwen3-ASR) model served by [vLLM](https://docs.vllm.ai/projects/recipes/en/latest/Qwen/Qwen3-ASR.html).
+
+This is not the Alibaba Model Studio realtime API. Point `base_url` at your own vLLM server.
+
+## Installation
+
+```bash
+pip install livekit-agents[qwen-asr]
+```
+
+## Run the model
+
+Batch transcription works with a normal serve:
+
+```bash
+vllm serve Qwen/Qwen3-ASR-1.7B
+```
+
+Realtime also needs the realtime architecture. That process still serves batch transcription, including context prompts:
+
+```bash
+vllm serve Qwen/Qwen3-ASR-1.7B \
+  --hf-overrides '{"architectures": ["Qwen3ASRRealtimeGeneration"]}'
+```
+
+## Batch
+
+`use_realtime=False` is the default. LiveKit's VAD cuts each turn and `recognize()` posts it to `/v1/audio/transcriptions`. Leave `prompt` unset for an empty context. Set it to bias names and domain words; vLLM puts that text in Qwen's system turn. LiveKit keyterms are appended to the same text, because the model has no separate keywords field.
+
+```python
+from livekit.plugins import qwen_asr
+
+stt = qwen_asr.STT(
+    base_url="http://127.0.0.1:8000/v1",
+    model="Qwen/Qwen3-ASR-1.7B",
+    language="tr",
+    prompt="Türkçe telefon görüşmesi. Fibabanka tüketici kredisi.",
+)
+```
+
+Omit `language` to let the model detect it. Omit `prompt` to send no context.
+
+## Realtime
+
+`use_realtime=True` opens vLLM's `/v1/realtime` socket. Audio is resampled to 16 kHz PCM. The server does not detect end of speech, so the plugin uses the bundled Silero VAD and closes the turn itself. Pass `vad=None` and call `flush()` to close turns yourself.
+
+```python
+stt = qwen_asr.STT(
+    base_url="http://127.0.0.1:8000/v1",
+    model="Qwen/Qwen3-ASR-1.7B",
+    language="tr",
+    use_realtime=True,
+)
+```
+
+`language` and `prompt` are sent on `session.update`. vLLM 0.30 reads the model name from that event and does not apply the prompt inside its realtime template, so context biasing currently applies to batch recognition. The plugin also drops the `language …<asr_text>` preamble that this vLLM version streams in front of the words.

livekit-plugins/livekit-plugins-qwen-asr/livekit/plugins/qwen_asr/__init__.py

@@ -0,0 +1,39 @@
+# Copyright 2026 LiveKit, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Qwen3-ASR plugin for a self-hosted vLLM server."""
+
+from livekit.agents import Plugin
+
+from .log import logger
+from .stt import STT, SpeechStream
+from .version import __version__
+
+__all__ = ["STT", "SpeechStream", "__version__"]
+
+
+class QwenAsrPlugin(Plugin):
+    def __init__(self) -> None:
+        super().__init__(__name__, __version__, __package__, logger)
+
+
+Plugin.register_plugin(QwenAsrPlugin())
+
+_module = dir()
+NOT_IN_ALL = [m for m in _module if m not in __all__]
+
+__pdoc__ = {}
+
+for n in NOT_IN_ALL:
+    __pdoc__[n] = False

livekit-plugins/livekit-plugins-qwen-asr/livekit/plugins/qwen_asr/log.py

@@ -0,0 +1,3 @@
+import logging
+
+logger = logging.getLogger("livekit.plugins.qwen_asr")