app+deps

README.md

@@ -1,7 +1,7 @@
 ---
-title: Asr Ukrainian
-emoji: ⚡
-colorFrom: red
+title: ASR Ukrainian
+emoji: 💬
+colorFrom: blue
 colorTo: yellow
 sdk: gradio
 app_file: app.py

app.py

@@ -0,0 +1,41 @@
+from transformers import pipeline, Wav2Vec2ProcessorWithLM
+from librosa import to_mono, resample
+import numpy as np
+import gradio as gr
+
+model_id = "arampacha/wav2vec2-xls-r-1b-uk"
+
+processor = Wav2Vec2ProcessorWithLM.from_pretrained(model_id)
+
+asr = pipeline(
+    "automatic-speech-recognition", model=model_id, device=-1, 
+    feature_extractor=processor.feature_extractor, decoder=processor.decoder
+)
+def run_asr(audio):
+    sr, audio_array = audio
+    audio_array = audio_array.astype(np.float32)
+    if len(audio_array.shape) > 1:
+        if audio_array.shape[1] == 1:
+            audio_array = audio_array.squeeze()
+        elif audio_array.shape[1] == 2:
+            audio_array = to_mono(audio_array.T)
+        else:
+            raise ValueError("Audio with > 2 channels not supported")
+    if sr != 16_000:
+        audio_array = resample(audio_array, orig_sr=sr, target_sr=16_000)
+    res = asr(audio_array, chunk_length_s=20, stride_length_s=2)
+    
+    return res["text"]
+
+text_out = gr.outputs.Textbox(label="transcript")
+interface = gr.Interface(
+    run_asr,
+    "microphone",
+    text_out,
+    layout="horizontal",
+    theme="huggingface",
+    title="Speech-to-text Ukrainian",
+    flagging_options=["incorrect"]
+)
+
+interface.launch()

packages.txt

@@ -0,0 +1 @@
+libsndfile1

requirements.txt

@@ -0,0 +1,6 @@
+-f https://download.pytorch.org/whl/cpu/torch_stable.html
+torch==1.10.2+cpu
+librosa
+transformers
+pypi-kenlm
+pyctcdecode