diff --git a/STT/paraformer_handler.py b/STT/paraformer_handler.py
index 0a2a9c002716b9acd1b3ddf1c52e847aa571ef2e..99fd6ac7912cc326472a31541a4ffdd9d8d79649 100644
--- a/STT/paraformer_handler.py
+++ b/STT/paraformer_handler.py
@@ -1,8 +1,6 @@
import logging
from time import perf_counter
-from tensorstore import dtype
-
from baseHandler import BaseHandler
from funasr import AutoModel
import numpy as np
@@ -19,22 +17,22 @@ console = Console()
class ParaformerSTTHandler(BaseHandler):
"""
- Handles the Speech To Text generation using a Whisper model.
+ Handles the Speech To Text generation using a Paraformer model.
+ The default for this model is set to Chinese.
+ This model was contributed by @wuhongsheng.
"""
def setup(
self,
model_name="paraformer-zh",
device="cuda",
- torch_dtype="float32",
- compile_mode=None,
gen_kwargs={},
):
print(model_name)
if len(model_name.split("/")) > 1:
model_name = model_name.split("/")[-1]
self.device = device
- self.model = AutoModel(model=model_name)
+ self.model = AutoModel(model=model_name, device=device)
self.warmup()
def warmup(self):
@@ -42,9 +40,9 @@ class ParaformerSTTHandler(BaseHandler):
# 2 warmup steps for no compile or compile mode with CUDA graphs capture
n_steps = 1
- dummy_input = np.array([0] * 512,dtype=np.float32)
+ dummy_input = np.array([0] * 512, dtype=np.float32)
for _ in range(n_steps):
- _ = self.model.generate(dummy_input)[0]["text"].strip().replace(" ","")
+ _ = self.model.generate(dummy_input)[0]["text"].strip().replace(" ", "")
def process(self, spoken_prompt):
logger.debug("infering paraformer...")
@@ -52,7 +50,9 @@ class ParaformerSTTHandler(BaseHandler):
global pipeline_start
pipeline_start = perf_counter()
- pred_text = self.model.generate(spoken_prompt)[0]["text"].strip().replace(" ","")
+ pred_text = (
+ self.model.generate(spoken_prompt)[0]["text"].strip().replace(" ", "")
+ )
torch.mps.empty_cache()
logger.debug("finished paraformer inference")
diff --git a/arguments_classes/module_arguments.py b/arguments_classes/module_arguments.py
index 140559641dcb9d908e1e426c89eb0815af318241..df9d94286965d23a75e2f71d5594f99aeb9148fe 100644
--- a/arguments_classes/module_arguments.py
+++ b/arguments_classes/module_arguments.py
@@ -23,7 +23,7 @@ class ModuleArguments:
stt: Optional[str] = field(
default="whisper",
metadata={
- "help": "The STT to use. Either 'whisper' or 'whisper-mlx'. Default is 'whisper'."
+ "help": "The STT to use. Either 'whisper', 'whisper-mlx', and 'paraformer'. Default is 'whisper'."
},
)
llm: Optional[str] = field(
diff --git a/arguments_classes/paraformer_stt_arguments.py b/arguments_classes/paraformer_stt_arguments.py
new file mode 100644
index 0000000000000000000000000000000000000000..a57a66abfbc6eb1f95868364a69e106919299032
--- /dev/null
+++ b/arguments_classes/paraformer_stt_arguments.py
@@ -0,0 +1,17 @@
+from dataclasses import dataclass, field
+
+
+@dataclass
+class ParaformerSTTHandlerArguments:
+ paraformer_stt_model_name: str = field(
+ default="paraformer-zh",
+ metadata={
+ "help": "The pretrained model to use. Default is 'paraformer-zh'. Can be choose from https://github.com/modelscope/FunASR"
+ },
+ )
+ paraformer_stt_device: str = field(
+ default="cuda",
+ metadata={
+ "help": "The device type on which the model will run. Default is 'cuda' for GPU acceleration."
+ },
+ )
diff --git a/requirements_mac.txt b/requirements_mac.txt
index e1e864a7da83171ac47b73175c5317a9614493bc..4897c33f635ec4f176313577dc60a8b4455f006f 100644
--- a/requirements_mac.txt
+++ b/requirements_mac.txt
@@ -5,5 +5,5 @@ torch==2.4.0
sounddevice==0.5.0
lightning-whisper-mlx>=0.0.10
mlx-lm>=0.14.0
-funasr
-modelscope
\ No newline at end of file
+funasr>=1.1.6
+modelscope>=1.17.1
\ No newline at end of file
diff --git a/s2s_pipeline.py b/s2s_pipeline.py
index a5e6aaa8cc3c142ca4838203ce417a2ffac90c41..1231abd824efe190d8036d78c7a508f058f6edb3 100644
--- a/s2s_pipeline.py
+++ b/s2s_pipeline.py
@@ -13,6 +13,7 @@ from arguments_classes.mlx_language_model_arguments import (
MLXLanguageModelHandlerArguments,
)
from arguments_classes.module_arguments import ModuleArguments
+from arguments_classes.paraformer_stt_arguments import ParaformerSTTHandlerArguments
from arguments_classes.parler_tts_arguments import ParlerTTSHandlerArguments
from arguments_classes.socket_receiver_arguments import SocketReceiverArguments
from arguments_classes.socket_sender_arguments import SocketSenderArguments
@@ -73,6 +74,7 @@ def main():
SocketSenderArguments,
VADHandlerArguments,
WhisperSTTHandlerArguments,
+ ParaformerSTTHandlerArguments,
LanguageModelHandlerArguments,
MLXLanguageModelHandlerArguments,
ParlerTTSHandlerArguments,
@@ -89,6 +91,7 @@ def main():
socket_sender_kwargs,
vad_handler_kwargs,
whisper_stt_handler_kwargs,
+ paraformer_stt_handler_kwargs,
language_model_handler_kwargs,
mlx_language_model_handler_kwargs,
parler_tts_handler_kwargs,
@@ -102,6 +105,7 @@ def main():
socket_sender_kwargs,
vad_handler_kwargs,
whisper_stt_handler_kwargs,
+ paraformer_stt_handler_kwargs,
language_model_handler_kwargs,
mlx_language_model_handler_kwargs,
parler_tts_handler_kwargs,
@@ -163,6 +167,8 @@ def main():
kwargs.tts_device = common_device
if hasattr(kwargs, "stt_device"):
kwargs.stt_device = common_device
+ if hasattr(kwargs, "paraformer_stt_device"):
+ kwargs.paraformer_stt_device = common_device
# Call this function with the common device and all the handlers
overwrite_device_argument(
@@ -171,9 +177,11 @@ def main():
mlx_language_model_handler_kwargs,
parler_tts_handler_kwargs,
whisper_stt_handler_kwargs,
+ paraformer_stt_handler_kwargs,
)
prepare_args(whisper_stt_handler_kwargs, "stt")
+ prepare_args(paraformer_stt_handler_kwargs, "paraformer_stt")
prepare_args(language_model_handler_kwargs, "lm")
prepare_args(mlx_language_model_handler_kwargs, "mlx_lm")
prepare_args(parler_tts_handler_kwargs, "tts")
@@ -245,14 +253,17 @@ def main():
)
elif module_kwargs.stt == "paraformer":
from STT.paraformer_handler import ParaformerSTTHandler
+
stt = ParaformerSTTHandler(
stop_event,
queue_in=spoken_prompt_queue,
queue_out=text_prompt_queue,
- # setup_kwargs=vars(whisper_stt_handler_kwargs),
+ setup_kwargs=vars(paraformer_stt_handler_kwargs),
)
else:
- raise ValueError("The STT should be either whisper or whisper-mlx")
+ raise ValueError(
+ "The STT should be either whisper, whisper-mlx, or paraformer."
+ )
if module_kwargs.llm == "transformers":
from LLM.language_model import LanguageModelHandler