commit 8c01ad5cd727b4844e26e8a050c4094b545c6938
parent 94a64876146c9ca4707c5d750dfe1e464db41735
Author: vin <git@vineetk.net>
Date: Wed, 13 Aug 2025 11:16:58 -0400
add parakeet asr via onnx
found out that nvidia's parakeet has onnx support.
it's about 4-5x faster than openai's whisper-large-v3-turbo (via
whisper.cpp) and seems more accurate too without lots of errors near the
end of the transcript.
I thought I lost this repo, turns out I wasn't stupid and already had it
pushed.
Diffstat:
| A | asr.py | | | 68 | ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ |
1 file changed, 68 insertions(+), 0 deletions(-)
diff --git a/asr.py b/asr.py
@@ -0,0 +1,68 @@
+import librosa
+import soundfile as sf
+import onnx_asr
+import os
+import sys
+#import numpy as np
+from tqdm import tqdm
+
+print("Loading ASR model...")
+providers = [
+ "ROCMExecutionProvider",
+ #"CUDAExecutionProvider",
+ "CPUExecutionProvider",
+]
+model = onnx_asr.load_model(
+ "nemo-parakeet-tdt-0.6b-v2",
+ providers=providers,
+ #quantization="int8"
+).with_timestamps()
+
+# parakeet needs 16khz mono wav as input
+input_file = "input.wav"
+chunk_duration_seconds = 20
+sample_rate = 16000
+
+print(f"Loading and resampling {input_file}...")
+try:
+ audio, sr = librosa.load(input_file, sr=sample_rate, mono=True)
+except Exception as e:
+ print(f"Error loading audio file: {e}")
+ exit()
+
+# process audio in chunks
+chunk_size_samples = int(chunk_duration_seconds * sample_rate)
+full_transcript = []
+temp_dir = "temp_chunks"
+os.makedirs(temp_dir, exist_ok=True)
+
+print("Starting transcription process in chunks...")
+num_chunks = (len(audio) + chunk_size_samples - 1) // chunk_size_samples
+
+for i in tqdm(range(num_chunks), desc="Transcribing"):
+ start_sample = i * chunk_size_samples
+ end_sample = start_sample + chunk_size_samples
+ chunk = audio[start_sample:end_sample]
+
+ temp_chunk_file = os.path.join(temp_dir, f"chunk_{i}.wav")
+ sf.write(temp_chunk_file, chunk, sample_rate)
+
+ try:
+ transcript = model.recognize(temp_chunk_file)
+ if transcript:
+ full_transcript.append(transcript)
+ except Exception as e:
+ print(f" Error processing chunk {i + 1}: {e}")
+ finally:
+ os.remove(temp_chunk_file)
+
+os.rmdir(temp_dir)
+
+print("\n" + "="*30)
+print(" FINAL TRANSCRIPT")
+print("="*30)
+print(full_transcript)
+
+# onnxruntime has some bug where it doesn't exit properly, aborting instead
+# so hard exit instead (sys.exit does graceful exit)
+os._exit(0)
