From 5043855ef829dd62943b0bf78ef89e80f4ae079b Mon Sep 17 00:00:00 2001
From: nikto_b <nikto_b@newt.team>
Date: Wed, 16 Sep 2026 00:07:29 +0300
Subject: [PATCH] 0.1.4.post3

---
 README.md                  |  2 +-
 pyproject.toml             | 20 ++++++++++++--------
 src/gigaam_onnx/asr_abc.py |  5 +++--
 src/gigaam_onnx/rnnt.py    |  6 +++++-
 4 files changed, 21 insertions(+), 12 deletions(-)

diff --git a/README.md b/README.md
index 11b2262..1666a67 100644
--- a/README.md
+++ b/README.md
@@ -58,7 +58,7 @@ import numpy as np
 if 'CUDAExecutionProvider' in ort.get_available_providers():
     provider = 'CUDAExecutionProvider'
 else:
-    provider = 'CPUExecutionProvider"
+    provider = 'CPUExecutionProvider'
 opts = ort.SessionOptions()
 opts.intra_op_num_threads = 16
 opts.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
diff --git a/pyproject.toml b/pyproject.toml
index aa6dc21..9bc42ba 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,22 +4,22 @@ authors = [
     { name = "nikto_b", email = "niktob560@yandex.ru" }
 ]
 license = "MIT"
-version = "0.1.3"
+version = "0.1.4.post3"
 description = "ONNX wrapper for a GigaAMASR models"
 readme = "README.md"
-requires-python = ">=3.10"
+requires-python = ">=3.8"
 dependencies = [
-    "hatchling>=1.28.0",
-    "numpy==2.*",
-    "onnxruntime==1.23.*",
-    "sentencepiece>=0.2.1",
-    "torch>=2.5,<2.9",
-    "torchaudio>=2.5,<2.9",
+    "numpy>=1.24.0",
+    "onnxruntime>=1.17.0",
+    "sentencepiece>=0.2.0",
+    "torch>=2.4,<2.9",
+    "torchaudio>=2.4,<2.9",
 ]
 
 [dependency-groups]
 dev = [
     "scipy>=1.0.0",
+    "hatchling",
 ]
 
 [build-system]
@@ -28,3 +28,7 @@ build-backend = "hatchling.build"
 
 [tool.hatch.build.targets.wheel]
 packages = ["src/gigaam_onnx"]
+core-metadata-version = "2.4"
+
+[tool.hatch.build.targets.sdist]
+core-metadata-version = "2.4"
diff --git a/src/gigaam_onnx/asr_abc.py b/src/gigaam_onnx/asr_abc.py
index 9281a87..1660ad8 100644
--- a/src/gigaam_onnx/asr_abc.py
+++ b/src/gigaam_onnx/asr_abc.py
@@ -76,9 +76,10 @@ class ASRABCModel(abc.ABC):
         features = None
         input_lengths_array = None
         full_batch_size = len(all_wavs)
+        BATCH_MAX_SIZE = 1
         while len(all_wavs) > 0:
-            wavs = all_wavs[:128]
-            all_wavs = all_wavs[128:]
+            wavs = all_wavs[:BATCH_MAX_SIZE]
+            all_wavs = all_wavs[BATCH_MAX_SIZE:]
 
             input_lengths = []
             processed_wavs = []
diff --git a/src/gigaam_onnx/rnnt.py b/src/gigaam_onnx/rnnt.py
index dadb5f9..4df7052 100644
--- a/src/gigaam_onnx/rnnt.py
+++ b/src/gigaam_onnx/rnnt.py
@@ -30,6 +30,9 @@ class RNNTASR(ASRABCModel):
         timings = []
         prev_token = self.blank_idx
 
+        if len(features.shape) == 2:
+            features = np.expand_dims(features, axis=0)
+
         pred_states = [
             np.zeros(shape=(1, 1, self.head.decoder.pred_hidden), dtype=DTYPE),
             np.zeros(shape=(1, 1, self.head.decoder.pred_hidden), dtype=DTYPE),
@@ -57,7 +60,8 @@ class RNNTASR(ASRABCModel):
                 log_probs = self.jointer.run(
                     [node.name for node in self.jointer.get_outputs()], joint_inputs
                 )
-                token = log_probs[0].argmax(-1)[0][0]
+                #token = log_probs[0].argmax(-1)[0][0]
+                token = int(log_probs[0].argmax(-1).reshape(-1)[0])
 
                 if token != self.blank_idx:
                     prev_token = int(token)