init

.gitignore

@@ -0,0 +1 @@
+.venv/

app.py

@@ -0,0 +1,137 @@
+import os
+import time
+
+import gradio as gr
+from pydub import AudioSegment
+from transformers import pipeline
+
+is_hf = os.getenv("SYSTEM") == "spaces"
+
+generate_kwargs = {
+    "language": "Japanese",
+    "do_sample": False,
+    "num_beams": 1,
+    "no_repeat_ngram_size": 3,
+}
+
+
+model_dict = {
+    "whisper-large-v2": "openai/whisper-large-v2",
+    "whisper-large-v3": "openai/whisper-large-v3",
+    "whisper-large-v3-turbo": "openai/whisper-large-v3-turbo",
+    "kotoba-whisper-v1.0": "kotoba-tech/kotoba-whisper-v1.0",
+    "kotoba-whisper-v2.0": "kotoba-tech/kotoba-whisper-v2.0",
+    "galgame-whisper-wip": (
+        "litagin/galgame-whisper-wip"
+        if is_hf
+        else "../whisper_finetune/galgame-whisper"
+    ),
+}
+
+# Download models
+for model in model_dict.values():
+    pipeline("automatic-speech-recognition", model=model)
+
+
+def transcribe_common(audio: str, model: str) -> tuple[str, float]:
+    # Get duration of audio
+    duration = AudioSegment.from_file(audio).duration_seconds
+    if duration > 15:
+        return "Audio too long, limit is 15 seconds", 0
+    start_time = time.time()
+    pipe = pipeline("automatic-speech-recognition", model=model)
+    end_time = time.time()
+    return pipe(audio, generate_kwargs=generate_kwargs)["text"], end_time - start_time
+
+
+def transcribe_large_v2(audio) -> tuple[str, float]:
+    return transcribe_common(audio, model_dict["whisper-large-v2"])
+
+
+def transcribe_large_v3(audio) -> tuple[str, float]:
+    return transcribe_common(audio, model_dict["whisper-large-v3"])
+
+
+def transcribe_large_v3_turbo(audio) -> tuple[str, float]:
+    return transcribe_common(audio, model_dict["whisper-large-v3-turbo"])
+
+
+def transcribe_kotoba_v1(audio) -> tuple[str, float]:
+    return transcribe_common(audio, model_dict["kotoba-whisper-v1.0"])
+
+
+def transcribe_kotoba_v2(audio) -> tuple[str, float]:
+    return transcribe_common(audio, model_dict["kotoba-whisper-v2.0"])
+
+
+def transcribe_galgame_whisper(audio) -> tuple[str, float]:
+    return transcribe_common(audio, model_dict["galgame-whisper-wip"])
+
+
+initial_md = """
+# Galgame-Whisper (WIP) Demo
+
+- 日本語のみ対応
+- 他の書き起こしとついでに比較できるようにいろいろ入れた
+- 現在0.1エポックくらい
+- 速度はCPUです
+- 音声は15秒まで
+"""
+
+with gr.Blocks() as app:
+    audio = gr.Audio(type="filepath")
+    gr.Markdown("### Kotoba-Whisper-V1.0")
+    with gr.Row():
+        with gr.Column():
+            gr.Markdown("### Whisper-Large-V2")
+            button_v2 = gr.Button("Transcribe with Whisper-Large-V2")
+            output_v2 = gr.Textbox()
+            time_v2 = gr.Textbox("Time taken")
+        with gr.Column():
+            gr.Markdown("### Whisper-Large-V3")
+            button_v3 = gr.Button("Transcribe with Whisper-Large-V3")
+            output_v3 = gr.Textbox()
+            time_v3 = gr.Textbox("Time taken")
+        with gr.Column():
+            gr.Markdown("### Whisper-Large-V3-Turbo")
+            button_v3_turbo = gr.Button("Transcribe with Whisper-Large-V3-Turbo")
+            output_v3_turbo = gr.Textbox()
+            time_v3_turbo = gr.Textbox()
+    with gr.Row():
+        with gr.Column():
+            gr.Markdown("### Kotoba-Whisper-V1.0")
+            button_kotoba_v1 = gr.Button("Transcribe with Kotoba-Whisper-V1.0")
+            output_kotoba_v1 = gr.Textbox()
+            time_kotoba_v1 = gr.Textbox("Time taken")
+        with gr.Column():
+            gr.Markdown("### Kotoba-Whisper-V2.0")
+            button_kotoba_v2 = gr.Button("Transcribe with Kotoba-Whisper-V2.0")
+            output_kotoba_v2 = gr.Textbox()
+            time_kotoba_v2 = gr.Textbox("Time taken")
+    with gr.Row():
+        with gr.Column():
+            gr.Markdown("### Galgame-Whisper (WIP)")
+            button_galgame = gr.Button("Transcribe with Galgame-Whisper (WIP)")
+            output_galgame = gr.Textbox()
+            time_galgame = gr.Textbox("Time taken")
+
+    button_v2.click(transcribe_large_v2, inputs=audio, outputs=[output_v2, time_v2])
+    button_v3.click(transcribe_large_v3, inputs=audio, outputs=[output_v3, time_v3])
+    button_v3_turbo.click(
+        transcribe_large_v3_turbo,
+        inputs=audio,
+        outputs=[output_v3_turbo, time_v3_turbo],
+    )
+    button_kotoba_v1.click(
+        transcribe_kotoba_v1, inputs=audio, outputs=[output_kotoba_v1, time_kotoba_v1]
+    )
+    button_kotoba_v2.click(
+        transcribe_kotoba_v2, inputs=audio, outputs=[output_kotoba_v2, time_kotoba_v2]
+    )
+    button_galgame.click(
+        transcribe_galgame_whisper,
+        inputs=audio,
+        outputs=[output_galgame, time_galgame],
+    )
+
+app.launch(inbrowser=True)

requirements.txt

@@ -0,0 +1,4 @@
+gradio
+numpy<2
+torch
+transformers