Update to version 0.54

README.md

@@ -21,13 +21,17 @@ model-index:
     metrics:
     - name: Test WER
       type: wer
-      value: 13.7
+      value: 11.3
 ---
 
 Small Zipformer2 model trained with k2-fsa/icefall on Russian data streaming version
 
+Version 0.54
+
 Links:
 
 <https://alphacephei.com/vosk>
 
 <https://github.com/k2-fsa/icefall>
+
+<https://github.com/k2-fsa/sherpa-onnx>

{am → am-onnx}/decoder.int8.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ef52d5fad1993813743f1ec86f2ebaef1ff9e083fbb214505af2066aefed3e39
-size 540292
+oid sha256:2b0df458692e1d090075c8249001136ef05240dd0d726a6b56552fd46c538b2d
+size 1326289

{am → am-onnx}/decoder.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7568cad4934d68114e2f19971d59d9272891f7ea5374ddbf52934dc31a9f9f0f
-size 2093079
+oid sha256:89b3088a9e20e1ef7f2e85ce1a3478afe6a9c4ac57369cabcc4beb8e95328ea0
+size 2093080

{am → am-onnx}/encoder.int8.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2e914e42ddf0c3601f1ab7ce930fd552e8a8b9fe038d1af6b254daec76c14221
-size 25817759
+oid sha256:e0db705e94ec35d803b1df4f40cda23d064e1142977c80ab288430b109777a9d
+size 26214060

{am → am-onnx}/encoder.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c7f925ed172032ab416eed7acdcd391185c1aba9742c3d312183f6c4b218f0a8
-size 91019510
+oid sha256:e9c27453e618bc97cf8a10169f34c104bd478166522907fcd122a46a88c78c69
+size 90994145

{am → am-onnx}/joiner.int8.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0a0288c8f04bff6a8abe3faf578ec5730a9bff1fc1fa2e2b914b1940f667a5f0
-size 259572
+oid sha256:b55784b071ab7512eab4c7c44e4f5478284ef33c83562cc6a249b972515a31e5
+size 259417

{am → am-onnx}/joiner.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7012e5520a3f196bf26a32ab7161eb2a6570ae3ab32654002ec6e1c1b6da0f72
-size 1026461
+oid sha256:dde0c7f3be0a16113a3e042c79a492c48667c07a8c1e9422ffe81c768aad4838
+size 1026462

am/epoch-32-avg-2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4ea2bee39c93481b3932e4cb4db326b0f91dc9dc0c87ef9bba54089b8f491ba
+size 380913148

decode.py

@@ -20,19 +20,21 @@ def read_wave(wave_filename: str) -> Tuple[np.ndarray, int]:
 def main():
 
     recognizer = sherpa_onnx.OnlineRecognizer.from_transducer(
-            encoder="am/encoder.onnx",
-            decoder="am/decoder.onnx",
-            joiner="am/joiner.onnx",
+            encoder="am-onnx/encoder.onnx",
+            decoder="am-onnx/decoder.onnx",
+            joiner="am-onnx/joiner.onnx",
             tokens="lang/tokens.txt",
             num_threads=4,
             sample_rate=16000,
-            decoding_method="greedy_search")
+            dither=3e-5,
+            decoding_method="modified_beam_search",
+            max_active_paths=10)
 
     samples, sample_rate = read_wave("test.wav")
 
     s = recognizer.create_stream()
     s.accept_waveform(sample_rate, waveform=samples)
-    tail_padding = np.zeros(int(sample_rate * 0.3)).astype(np.float32)
+    tail_padding = np.zeros(int(sample_rate * 0.6)).astype(np.float32)
     s.accept_waveform(sample_rate, waveform=tail_padding)
     s.input_finished()
 

decode8.py

@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+import wave
+from pathlib import Path
+from typing import Tuple
+import sys
+import numpy as np
+import sherpa_onnx
+
+def read_wave(wave_filename: str) -> Tuple[np.ndarray, int]:
+    with wave.open(wave_filename) as f:
+        assert f.getnchannels() == 1, f.getnchannels()
+        assert f.getsampwidth() == 2, f.getsampwidth()  # it is in bytes
+        num_samples = f.getnframes()
+        samples = f.readframes(num_samples)
+        samples_int16 = np.frombuffer(samples, dtype=np.int16)
+        samples_float32 = samples_int16.astype(np.float32)
+        samples_float32 = samples_float32 / 32768
+        return samples_float32, f.getframerate()
+
+def main():
+
+    recognizer = sherpa_onnx.OnlineRecognizer.from_transducer(
+            encoder="am-onnx/encoder.int8.onnx",
+            decoder="am-onnx/decoder.int8.onnx",
+            joiner="am-onnx/joiner.int8.onnx",
+            tokens="lang/tokens.txt",
+            num_threads=4,
+            sample_rate=16000,
+            dither=3e-5,
+            decoding_method="modified_beam_search",
+            max_active_paths=10)
+
+    samples, sample_rate = read_wave("test.wav")
+
+    s = recognizer.create_stream()
+    s.accept_waveform(sample_rate, waveform=samples)
+    tail_padding = np.zeros(int(sample_rate * 0.6)).astype(np.float32)
+    s.accept_waveform(sample_rate, waveform=tail_padding)
+    s.input_finished()
+
+    while recognizer.is_ready(s):
+        recognizer.decode_stream(s)
+        print (recognizer.get_result(s))
+
+if __name__ == "__main__":
+    main()
+