Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

Fix Fun-ASR-Nano stereo audio preprocessing (#3102)

zhifu gaocb7adc018428
diff --git a/examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py b/examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py
index 99e89f759..feae0c6ef 100644
--- a/examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py
+++ b/examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py
@@ -67,6 +67,21 @@ _spk_model = None
_args = None
+def prepare_audio_for_inference(audio_data, sr, target_sr=16000):
+ """Return mono float32 audio at target_sr for ASR inference."""
+ audio_data = np.asarray(audio_data)
+ if audio_data.ndim > 1:
+ channel_axis = -1 if audio_data.shape[-1] <= audio_data.shape[0] else 0
+ audio_data = audio_data.mean(axis=channel_axis)
+
+ if sr != target_sr:
+ import librosa
+ audio_data = librosa.resample(audio_data, orig_sr=sr, target_sr=target_sr)
+ sr = target_sr
+
+ return audio_data.astype(np.float32), sr
+
+
def load_engine(args):
global _engine, _vad_model, _spk_model, _args
_args = args
@@ -90,14 +105,7 @@ def load_engine(args):
def process_audio(audio_data, sr=16000, language=None, hotwords=None,
use_vad=True, use_spk=False, use_timestamp=True):
"""Core processing: VAD segment → vLLM ASR → timestamps → SPK."""
- if sr != 16000:
- import librosa
- audio_data = librosa.resample(audio_data, orig_sr=sr, target_sr=16000)
- sr = 16000
-
- if audio_data.ndim > 1:
- audio_data = audio_data[:, 0]
- audio_data = audio_data.astype(np.float32)
+ audio_data, sr = prepare_audio_for_inference(audio_data, sr)
# VAD segmentation
if use_vad and len(audio_data) > sr * 1:
diff --git a/funasr/bin/_server_app.py b/funasr/bin/_server_app.py
index 17c9ba676..90ff45580 100644
--- a/funasr/bin/_server_app.py
+++ b/funasr/bin/_server_app.py
@@ -26,6 +26,20 @@ except ImportError:
logger = logging.getLogger("funasr.server")
+def prepare_audio_for_inference(audio_data, sr, target_sr=16000):
+ """Return mono float32 audio at target_sr for ASR inference."""
+ audio_data = np.asarray(audio_data)
+ if audio_data.ndim > 1:
+ channel_axis = -1 if audio_data.shape[-1] <= audio_data.shape[0] else 0
+ audio_data = audio_data.mean(axis=channel_axis)
+
+ if sr != target_sr:
+ import librosa
+ audio_data = librosa.resample(audio_data, orig_sr=sr, target_sr=target_sr)
+ sr = target_sr
+
+ return audio_data.astype(np.float32), sr
+
def create_app(device: str = "cuda", preload_model: str = "auto") -> FastAPI:
if preload_model == "auto":
preload_model = "fun-asr-nano" if device.startswith("cuda") else "sensevoice"
@@ -107,13 +121,7 @@ def create_app(device: str = "cuda", preload_model: str = "auto") -> FastAPI:
def _process_vllm(audio_data, sr, language=None, hotwords=None, use_spk=False):
"""Process audio with vLLM engine (Fun-ASR-Nano)."""
- if sr != 16000:
- import librosa
- audio_data = librosa.resample(audio_data, orig_sr=sr, target_sr=16000)
- sr = 16000
- if audio_data.ndim > 1:
- audio_data = audio_data[:, 0]
- audio_data = audio_data.astype(np.float32)
+ audio_data, sr = prepare_audio_for_inference(audio_data, sr)
# VAD
vad_res = app.state.vad_model.generate(input=audio_data, fs=sr)
diff --git a/tests/test_fun_asr_nano_openai_response.py b/tests/test_fun_asr_nano_openai_response.py
index 4b836b80d..9f04eb81c 100644
--- a/tests/test_fun_asr_nano_openai_response.py
+++ b/tests/test_fun_asr_nano_openai_response.py
@@ -3,6 +3,8 @@ import sys
import types
from pathlib import Path
+import numpy as np
+
REPO_ROOT = Path(__file__).resolve().parents[1]
SERVICE_PATH = (
@@ -108,3 +110,44 @@ def test_openai_verbose_json_keeps_segment_speaker(monkeypatch):
)
assert response["segments"][0]["speaker"] == "SPK0"
+
+
+def test_process_audio_downmixes_stereo_before_resampling(monkeypatch):
+ module = load_service_module(monkeypatch)
+ captured = {}
+
+ librosa_stub = types.ModuleType("librosa")
+
+ def fake_resample(audio, orig_sr, target_sr):
+ assert audio.ndim == 1
+ captured["resample_input"] = audio.copy()
+ assert orig_sr == 8000
+ assert target_sr == 16000
+ return np.repeat(audio, 2)
+
+ librosa_stub.resample = fake_resample
+ monkeypatch.setitem(sys.modules, "librosa", librosa_stub)
+
+ class EngineStub:
+ def generate(self, inputs, **kwargs):
+ captured["engine_input"] = inputs[0]
+ return [{"text": "ok"}]
+
+ module._engine = EngineStub()
+ stereo_audio = np.column_stack(
+ [np.zeros(8000, dtype=np.float32), np.ones(8000, dtype=np.float32)]
+ )
+
+ result = module.process_audio(
+ stereo_audio,
+ sr=8000,
+ use_vad=False,
+ use_spk=False,
+ use_timestamp=False,
+ )
+
+ assert np.allclose(captured["resample_input"], 0.5)
+ assert captured["engine_input"].shape == (16000,)
+ assert np.allclose(captured["engine_input"], 0.5)
+ assert result["duration"] == 1.0
+ assert result["text"] == "ok"