Python
This page shows how to use X-ASR models with sherpa-onnx in Python. Two demos are given: one with a non-streaming model and one with a streaming model.
On this page
Requirements
Python 3.8 or above
sherpa-onnx, installed with
pip install sherpa-onnxsoundfile:pip install soundfilenumpy, which is installed together with sherpa-onnx
Non-streaming X-ASR
Step 1. Create the project directory
mkdir x-asr-demo
cd x-asr-demo
Step 2. Download the model
wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-int8-2026-06-03.tar.bz2
tar xf sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-int8-2026-06-03.tar.bz2
rm sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-int8-2026-06-03.tar.bz2
Step 3. Save the code
Save the following as x-asr-offline.py:
#!/usr/bin/env python3
# This file shows how to use a non-streaming X-ASR model with sherpa-onnx
# in Python.
#
# Before you run it, please download the model:
#
# wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-int8-2026-06-03.tar.bz2
# tar xf sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-int8-2026-06-03.tar.bz2
#
# and install sherpa-onnx:
#
# pip install sherpa-onnx
import time
import numpy as np
import soundfile as sf
import sherpa_onnx
def read_wave(filename: str):
"""Read a WAVE file.
Returns:
A tuple of (samples, sample_rate), where samples is a 1-D float32
array normalized to the range [-1, 1].
"""
# See https://python-soundfile.readthedocs.io/en/0.11.0/#soundfile.read
data, sample_rate = sf.read(
filename,
always_2d=True,
dtype="float32",
)
data = data[:, 0] # use only the first channel
samples = np.ascontiguousarray(data)
# samples is a 1-D array of dtype float32
# sample_rate is a scalar
return samples, sample_rate
def main():
model_dir = "./sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-int8-2026-06-03"
# Step 1: Create the recognizer from the X-ASR model files.
recognizer = sherpa_onnx.OfflineRecognizer.from_transducer(
encoder=f"{model_dir}/encoder-epoch-99-avg-1.int8.onnx",
decoder=f"{model_dir}/decoder-epoch-99-avg-1.onnx",
joiner=f"{model_dir}/joiner-epoch-99-avg-1.int8.onnx",
tokens=f"{model_dir}/tokens.txt",
num_threads=2,
debug=False,
)
# Step 2: Read the wave file you want to transcribe.
filename = f"{model_dir}/test_wavs/0.wav"
samples, sample_rate = read_wave(filename)
# Step 3: Create a stream, send the whole audio to it, and decode.
start = time.time()
stream = recognizer.create_stream()
stream.accept_waveform(sample_rate, samples)
recognizer.decode_stream(stream)
elapsed = time.time() - start
# Step 4: Print the result. X-ASR punct models also add punctuations.
print(f"{filename}\n{stream.result.text}")
audio_duration = len(samples) / sample_rate
print(f"audio duration: {audio_duration:.3f} s")
print(f"elapsed seconds: {elapsed:.3f} s")
print(f"RTF = {elapsed:.3f}/{audio_duration:.3f} = {elapsed / audio_duration:.3f}")
if __name__ == "__main__":
main()
Step 4. Check the project layout
Your project directory should now look like this:
$ tree .
.
├── sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-int8-2026-06-03
│ ├── bpe.model
│ ├── decoder-epoch-99-avg-1.onnx
│ ├── encoder-epoch-99-avg-1.int8.onnx
│ ├── joiner-epoch-99-avg-1.int8.onnx
│ ├── README.md
│ ├── test_onnx.py
│ ├── test_wavs
│ │ ├── 0.wav
│ │ ├── 1.wav
│ │ ├── 2.wav
│ │ └── 3.wav
│ └── tokens.txt
└── x-asr-offline.py
3 directories, 12 files
$ ls -lh
total 8
drwxr-xr-x@ 10 fangjun wheel 320B 29 Jun 16:31 sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-int8-2026-06-03
-rw-r--r--@ 1 fangjun wheel 2.4K 19 Nov 12:23 x-asr-offline.py
Step 5. Build and run
python3 ./x-asr-offline.py
Output
./sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-int8-2026-06-03/test_wavs/0.wav
昨天是 Monday, today is 礼拜二, the day after tomorrow 是星期三。
audio duration: 10.053 s
elapsed seconds: 0.129 s
RTF = 0.129/10.053 = 0.013
Note
The timing above was measured on a MacBook Pro (Apple M4) using 2 threads. Please increase the number of threads for a smaller RTF on your machine.
Streaming X-ASR
Step 1. Create the project directory
mkdir x-asr-demo
cd x-asr-demo
Step 2. Download the model
wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-x-asr-480ms-streaming-zipformer-transducer-zh-en-punct-int8-2026-06-05.tar.bz2
tar xf sherpa-onnx-x-asr-480ms-streaming-zipformer-transducer-zh-en-punct-int8-2026-06-05.tar.bz2
rm sherpa-onnx-x-asr-480ms-streaming-zipformer-transducer-zh-en-punct-int8-2026-06-05.tar.bz2
Step 3. Save the code
Save the following as x-asr-streaming.py:
#!/usr/bin/env python3
# This file shows how to use a streaming X-ASR model with sherpa-onnx
# in Python.
#
# Before you run it, please download the model:
#
# wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-x-asr-480ms-streaming-zipformer-transducer-zh-en-punct-int8-2026-06-05.tar.bz2
# tar xf sherpa-onnx-x-asr-480ms-streaming-zipformer-transducer-zh-en-punct-int8-2026-06-05.tar.bz2
#
# and install sherpa-onnx:
#
# pip install sherpa-onnx
import time
import numpy as np
import soundfile as sf
import sherpa_onnx
def read_wave(filename: str):
"""Read a WAVE file.
Returns:
A tuple of (samples, sample_rate), where samples is a 1-D float32
array normalized to the range [-1, 1].
"""
# See https://python-soundfile.readthedocs.io/en/0.11.0/#soundfile.read
data, sample_rate = sf.read(
filename,
always_2d=True,
dtype="float32",
)
data = data[:, 0] # use only the first channel
samples = np.ascontiguousarray(data)
# samples is a 1-D array of dtype float32
# sample_rate is a scalar
return samples, sample_rate
def main():
model_dir = (
"./sherpa-onnx-x-asr-480ms-streaming-zipformer-transducer-zh-en-punct-int8-2026-06-05"
)
# Step 1: Create the recognizer from the X-ASR model files.
recognizer = sherpa_onnx.OnlineRecognizer.from_transducer(
encoder=f"{model_dir}/encoder.int8.onnx",
decoder=f"{model_dir}/decoder.onnx",
joiner=f"{model_dir}/joiner.int8.onnx",
tokens=f"{model_dir}/tokens.txt",
num_threads=2,
debug=False,
)
# Step 2: Read the wave file you want to transcribe.
filename = f"{model_dir}/test_wavs/0.wav"
samples, sample_rate = read_wave(filename)
# Step 3: Create a stream and feed the audio in small chunks, just like
# a microphone would do in a real-time application.
start = time.time()
stream = recognizer.create_stream()
# 0.2 second for 16 kHz audio
chunk_size = int(0.2 * sample_rate)
last = ""
for i in range(0, len(samples), chunk_size):
stream.accept_waveform(sample_rate, samples[i : i + chunk_size])
while recognizer.is_ready(stream):
recognizer.decode_stream(stream)
# Print the latest partial result when it changes.
result = recognizer.get_result(stream)
if result and result != last:
last = result
print(result)
# Send some tail padding so the last words are flushed out.
stream.accept_waveform(sample_rate, np.zeros(int(0.5 * sample_rate), dtype=np.float32))
while recognizer.is_ready(stream):
recognizer.decode_stream(stream)
# Step 4: Print the final result.
elapsed = time.time() - start
print("Final result:", recognizer.get_result(stream))
audio_duration = len(samples) / sample_rate
print(f"audio duration: {audio_duration:.3f} s")
print(f"elapsed seconds: {elapsed:.3f} s")
print(f"RTF = {elapsed:.3f}/{audio_duration:.3f} = {elapsed / audio_duration:.3f}")
if __name__ == "__main__":
main()
Step 4. Check the project layout
Your project directory should now look like this:
$ tree .
.
├── sherpa-onnx-x-asr-480ms-streaming-zipformer-transducer-zh-en-punct-int8-2026-06-05
│ ├── bpe.model
│ ├── decoder.onnx
│ ├── encoder.int8.onnx
│ ├── joiner.int8.onnx
│ ├── README.md
│ ├── test_onnx.py
│ ├── test_wavs
│ │ ├── 0.wav
│ │ ├── 1.wav
│ │ ├── 2.wav
│ │ └── 3.wav
│ └── tokens.txt
└── x-asr-streaming.py
3 directories, 12 files
$ ls -lh
total 8
drwxr-xr-x@ 10 fangjun wheel 320B 29 Jun 16:30 sherpa-onnx-x-asr-480ms-streaming-zipformer-transducer-zh-en-punct-int8-2026-06-05
-rw-r--r--@ 1 fangjun wheel 3.1K 19 Nov 12:23 x-asr-streaming.py
Step 5. Build and run
python3 ./x-asr-streaming.py
Output
昨
昨天是
昨天是 Mon
昨天是 Monday
昨天是 Monday,
昨天是 Monday, today
昨天是 Monday, today is 礼
昨天是 Monday, today is 礼拜二
昨天是 Monday, today is 礼拜二,
昨天是 Monday, today is 礼拜二, the day
昨天是 Monday, today is 礼拜二, the day after
昨天是 Monday, today is 礼拜二, the day after tomorro
昨天是 Monday, today is 礼拜二, the day after tomorrow
昨天是 Monday, today is 礼拜二, the day after tomorrow 是
Final result: 昨天是 Monday, today is 礼拜二, the day after tomorrow 是星期
audio duration: 10.053 s
elapsed seconds: 0.305 s
RTF = 0.305/10.053 = 0.030
Note
The timing above was measured on a MacBook Pro (Apple M4) using 2 threads. Please increase the number of threads for a smaller RTF on your machine.