From 2a5ef6997772184e6b2f344762134d90a7ad8037 Mon Sep 17 00:00:00 2001
From: Azalea Gui <22280294+hykilpikonna@users.noreply.github.com>
Date: Sat, 7 Dec 2024 10:23:04 -0500
Subject: [PATCH] [+] Transcribe audio with faster-whisper

---
 scripts/bin/transcribe | 64 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 64 insertions(+)
 create mode 100755 scripts/bin/transcribe

diff --git a/scripts/bin/transcribe b/scripts/bin/transcribe
new file mode 100755
index 0000000..c9c5473
--- /dev/null
+++ b/scripts/bin/transcribe
@@ -0,0 +1,64 @@
+#!/usr/bin/env python3
+import argparse
+from pathlib import Path
+from faster_whisper import WhisperModel, BatchedInferencePipeline
+# import nemo.collections.asr as nemo_asr
+# asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained(model_name="nvidia/parakeet-tdt-1.1b")
+# import nemo.collections.asr.models.rnnt_bpe_models.EncDecRNNTBPEModel
+# asr_model.transcribe
+
+
+# model_name = 'deepdml/faster-whisper-large-v3-turbo-ct2'
+model_name = 'distil-large-v3'
+# model_name = 'medium.en'
+m = WhisperModel(model_name, device="cuda", compute_type="float16")
+model = BatchedInferencePipeline(model=m)
+
+
+def format_time(seconds):
+    minutes, seconds = divmod(seconds, 60)
+    hours, minutes = divmod(minutes, 60)
+    milliseconds = (seconds - int(seconds)) * 1000
+    return f"{int(hours):02d}:{int(minutes):02d}:{int(seconds):02d},{int(milliseconds):03d}"
+
+
+def transcribe(input_file: Path, lang: str):
+    ouf = input_file.with_suffix('.srt')
+    if ouf.exists():
+        print(f"Output file {ouf} already exists. Skipping transcription.")
+        return
+
+    # Remove task="translate" if you want the original language
+    segments, info = model.transcribe(input_file, beam_size=1, batch_size=8, 
+                                    #   chunk_length=10, 
+                                      without_timestamps=False,
+                                      task="transcribe", vad_filter=True, language=lang)
+
+    print(f"Transcribing file {input_file}")
+    print(f"Detected language '{info.language}' with probability {info.language_probability:.2f}")
+
+    # with ouf.open('w', encoding='utf-8') as srt_file:
+    out = ""
+    for seg in segments:
+        start_time = format_time(seg.start)
+        end_time = format_time(seg.end)
+        line_out = f"{seg.id + 1}\n{start_time} --> {end_time}\n{seg.text.lstrip()}\n\n"
+        print(line_out)
+        out += line_out
+    
+    ouf.write_text(out)
+    print(f"Transcription saved to {ouf}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Transcribe audio from a video file and generate an SRT file.")
+    # parser.add_argument("input_file", help="Path to the video file for transcription")
+    parser.add_argument("input_file", nargs="+", help="Path to the video file for transcription")
+    parser.add_argument("-l", "--lang", default="en", help="Language code for transcription")
+    args = parser.parse_args()
+    for file in args.input_file:
+        transcribe(Path(file), args.lang)
+
+
+if __name__ == "__main__":
+    main()