Add gradio application with scorer

README.md

@@ -1,3 +1,13 @@
+---
+title: "Ukrainian Speech-to-Text"
+emoji: 🐸
+colorFrom: blue
+colorTo: yellow
+sdk: gradio
+app_file: app.py
+pinned: false
+---
+
 # voice-recognition-ua
 This is a repository with aim to apply [Coqui STT](https://github.com/coqui-ai/STT "STT")(formerly [DeepSpeech](https://github.com/mozilla/DeepSpeech)) (state-of-the-art speech recognition model) on Ukrainian language.  
 You can see online demo here: https://voice-recognition-ua.herokuapp.com (your voice is not stored).  

app.py

@@ -0,0 +1,102 @@
+from io import BytesIO
+from typing import Tuple
+import wave
+import gradio as gr
+import numpy as np
+from pydub.audio_segment import AudioSegment
+import requests
+from os.path import exists
+from stt import Model
+
+
+MODEL_NAMES = [
+    "With scorer",
+    "No scorer"
+]
+
+# download model
+version = "v0.4"
+storage_url = f"https://github.com/robinhad/voice-recognition-ua/releases/download/{version}"
+model_name = "uk.tflite"
+scorer_name = "kenlm.scorer"
+model_link = f"{storage_url}/{model_name}"
+scorer_link = f"{storage_url}/{scorer_name}"
+
+
+def client(audio_data: np.array, sample_rate: int, use_scorer=False):
+    output_audio = _convert_audio(audio_data, sample_rate)
+
+    fin = wave.open(output_audio, 'rb')
+    audio = np.frombuffer(fin.readframes(fin.getnframes()), np.int16)
+
+    fin.close()
+
+    ds = Model(model_name)
+    if use_scorer:
+        ds.enableExternalScorer("kenlm.scorer")
+
+    result = ds.stt(audio)
+
+    return result
+
+
+def download(url, file_name):
+    if not exists(file_name):
+        print(f"Downloading {file_name}")
+        r = requests.get(url, allow_redirects=True)
+        with open(file_name, 'wb') as file:
+            file.write(r.content)
+    else:
+        print(f"Found {file_name}. Skipping download...")
+
+
+def stt(audio: Tuple[int, np.array], model_name: str):
+    sample_rate, audio = audio
+    use_scorer = True if model_name == "With scorer" else False
+
+    if sample_rate != 16000:
+        raise ValueError("Incorrect sample rate.")
+
+    recognized_result = client(audio, sample_rate, use_scorer)
+
+    return recognized_result
+
+
+def _convert_audio(audio_data: np.array, sample_rate: int):
+    source_audio = BytesIO()
+    source_audio.write(audio_data)
+    source_audio.seek(0)
+    output_audio = BytesIO()
+    wav_file = AudioSegment.from_raw(
+        source_audio,
+        channels=1,
+        sample_width=2,
+        frame_rate=sample_rate
+    )
+    wav_file.set_frame_rate(16000).set_channels(
+        1).export(output_audio, "wav", codec="pcm_s16le")
+    output_audio.seek(0)
+    return output_audio
+
+
+iface = gr.Interface(
+    fn=stt,
+    inputs=[
+        gr.inputs.Audio(type="numpy",
+                        label=None, optional=False),
+        gr.inputs.Radio(
+            label="Виберіть Speech-to-Text модель",
+            choices=MODEL_NAMES,
+        ),
+
+    ],
+    outputs=gr.outputs.Textbox(label="Output"),
+    title="🐸🇺🇦 - Coqui STT",
+    theme="huggingface",
+    description="Україномовний🇺🇦 Speech-to-Text за допомогою Coqui STT",
+    article="Якщо вам подобається, підтримайте за посиланням: [SUPPORT LINK](https://send.monobank.ua/jar/48iHq4xAXm)",
+)
+
+download(model_link, model_name)
+download(scorer_link, scorer_name)
+iface.launch()

packages.txt

@@ -0,0 +1 @@
+ffmpeg

requirements.txt

@@ -0,0 +1,3 @@
+gradio==2.4.5
+STT==1.0.0
+pydub==0.25.1