77 lines
2.4 KiB
Python
77 lines
2.4 KiB
Python
#! /usr/bin/env python
|
|
# pip install -r requirements.txt
|
|
usage = 'python transcribe-voice-track.py <wav filenames...> '
|
|
|
|
# https://towardsdatascience.com/speech-recognition-with-timestamps-934ede4234b2
|
|
# If you don't get results, try re-exporting as Signed 16-bit PCM
|
|
|
|
import util
|
|
import os
|
|
import requests
|
|
from zipfile import ZipFile
|
|
|
|
from vosk import Model, KaldiRecognizer
|
|
|
|
model_name = "vosk-model-small-en-us-0.15"
|
|
# model_name = "vosk-model-en-us-0.22" # bigger model. So far, the small one works pretty good.
|
|
|
|
model_path = f"models/{model_name}"
|
|
model_zip_path = f"{model_path}.zip"
|
|
model_url = f"http://alphacephei.com/vosk/{model_zip_path}"
|
|
|
|
# Download the model if it doesn't exist
|
|
os.makedirs('models', exist_ok=True)
|
|
if not os.path.exists(model_path):
|
|
with open(model_zip_path, 'wb') as f:
|
|
response = requests.get(model_url)
|
|
f.write(response.content)
|
|
with ZipFile(model_zip_path, "r") as zip_file:
|
|
zip_file.extractall('models')
|
|
|
|
model = Model(model_path)
|
|
|
|
audio_filenames = util.args(1, usage)
|
|
for audio_filename in audio_filenames:
|
|
wf = wave.open(audio_filename, "rb")
|
|
|
|
rec = KaldiRecognizer(model, wf.getframerate())
|
|
rec.SetWords(True)
|
|
|
|
frames = 4000
|
|
|
|
# Mix channels together if the input is stereo
|
|
# or the sample width is incompatible
|
|
if wf.getnchannels() == 2:
|
|
wf.close()
|
|
mono_filename = '.'.join(audio_filename.split('.')[:-1]) + '_mono.wav'
|
|
fs, data = wavfile.read(audio_filename)
|
|
wavfile.write(mono_filename, fs, data[:, 0])
|
|
wf = wave.open(mono_filename, 'rb')
|
|
|
|
wf.rewind()
|
|
results = []
|
|
def filter_result(result):
|
|
result = json.loads(result)
|
|
if len(result) != 1:
|
|
results.append(result)
|
|
|
|
while True:
|
|
data = wf.readframes(frames)
|
|
if len(data) == 0:
|
|
break
|
|
if rec.AcceptWaveform(data):
|
|
filter_result(rec.Result())
|
|
filter_result(rec.FinalResult())
|
|
|
|
with open(f"{audio_filename}_{frames}.json", "w") as f:
|
|
lines = {}
|
|
for sentence in results:
|
|
words = sentence['result']
|
|
text = sentence['text']
|
|
# Account for duplicate sentences:
|
|
if not text in lines:
|
|
lines[text] = []
|
|
lines[text].append({'start': words[0]['start'], 'end': words[-1]['end']})
|
|
print(f'{text}: {words[0]["start"]} {words[-1]["end"]}')
|
|
json.dump(lines, f)
|