malarinv/plume-asr
1
0
Fork 0
mirror of https://github.com/malarinv/plume-asr.git synced 2026-03-08 04:12:35 +00:00
Code Issues Releases Wiki Activity

massive refactor/rename to plume

Browse Source
This commit is contained in:
Malar Kannan
2021-02-23 19:43:33 +05:30
parent e8f58a5043
commit ed6117559a
51 changed files with 2864 additions and 1037 deletions
Download Patch File Download Diff File Expand all files Collapse all files

486
plume/utils/__init__.py Normal file
View File

@@ -0,0 +1,486 @@
import io
import os
import re
import json
import wave
import logging
from pathlib import Path
from functools import partial
from uuid import uuid4
from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor
import subprocess
import shutil
from urllib.parse import urlsplit
# from .lazy_loader import LazyLoader
from .lazy_import import lazy_callable, lazy_module
# from ruamel.yaml import YAML
# import boto3
import typer
# import pymongo
# from slugify import slugify
# import pydub
# import matplotlib.pyplot as plt
# import librosa
# import librosa.display as audio_display
# from natural.date import compress
# from num2words import num2words
from tqdm import tqdm
from datetime import timedelta
# from .transcribe import triton_transcribe_grpc_gen
# from .eval import app as eval_app
from .tts import app as tts_app
from .transcribe import app as transcribe_app
from .align import app as align_app
boto3 = lazy_module('boto3')
pymongo = lazy_module('pymongo')
pydub = lazy_module('pydub')
audio_display = lazy_module('librosa.display')
plt = lazy_module('matplotlib.pyplot')
librosa = lazy_module('librosa')
YAML = lazy_callable('ruamel.yaml.YAML')
num2words = lazy_callable('num2words.num2words')
slugify = lazy_callable('slugify.slugify')
compress = lazy_callable('natural.date.compress')
app = typer.Typer()
app.add_typer(tts_app, name="tts")
app.add_typer(align_app, name="align")
app.add_typer(transcribe_app, name="transcribe")
logging.basicConfig(
level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s"
)
logger = logging.getLogger(__name__)
def manifest_str(path, dur, text):
return (
json.dumps({"audio_filepath": path, "duration": round(dur, 1), "text": text})
+ "\n"
)
def duration_str(seconds):
return compress(timedelta(seconds=seconds), pad=" ")
def replace_digit_symbol(w2v_out):
num_int_map = {num2words(i): str(i) for i in range(10)}
out = w2v_out.lower()
for (k, v) in num_int_map.items():
out = re.sub(k, v, out)
return out
def discard_except_digits(inp):
return re.sub("[^0-9]", "", inp)
def digits_to_chars(text):
num_tokens = [num2words(c) + " " if "0" <= c <= "9" else c for c in text]
return ("".join(num_tokens)).lower()
def replace_redundant_spaces_with(text, sub):
return re.sub(" +", sub, text)
def space_out(text):
letters = " ".join(list(text))
return letters
def wav_bytes(audio_bytes, frame_rate=24000):
wf_b = io.BytesIO()
with wave.open(wf_b, mode="w") as wf:
wf.setnchannels(1)
wf.setframerate(frame_rate)
wf.setsampwidth(2)
wf.writeframesraw(audio_bytes)
return wf_b.getvalue()
def tscript_uuid_fname(transcript):
return str(uuid4()) + "_" + slugify(transcript, max_length=8)
def run_shell(cmd_str, work_dir="."):
cwd_path = Path(work_dir).absolute()
p = subprocess.Popen(
cmd_str,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
shell=True,
cwd=cwd_path,
)
for line in p.stdout:
print(line.replace(b"\n", b"").decode("utf-8"))
def upload_s3(dataset_path, s3_path):
run_shell(f"aws s3 sync {dataset_path} {s3_path}")
def get_download_path(s3_uri, output_path):
s3_uri_p = urlsplit(s3_uri)
download_path = output_path / Path(s3_uri_p.path[1:])
download_path.parent.mkdir(exist_ok=True, parents=True)
return download_path
def s3_downloader():
s3 = boto3.client("s3")
def download_s3(s3_uri, download_path):
s3_uri_p = urlsplit(s3_uri)
download_path.parent.mkdir(exist_ok=True, parents=True)
if not download_path.exists():
print(f"downloading {s3_uri} to {download_path}")
s3.download_file(s3_uri_p.netloc, s3_uri_p.path[1:], str(download_path))
return download_s3
def asr_data_writer(dataset_dir, asr_data_source, verbose=False):
(dataset_dir / Path("wavs")).mkdir(parents=True, exist_ok=True)
asr_manifest = dataset_dir / Path("manifest.json")
num_datapoints = 0
with asr_manifest.open("w") as mf:
print(f"writing manifest to {asr_manifest}")
for transcript, audio_dur, wav_data in asr_data_source:
fname = tscript_uuid_fname(transcript)
audio_file = dataset_dir / Path("wavs") / Path(fname).with_suffix(".wav")
audio_file.write_bytes(wav_data)
rel_data_path = audio_file.relative_to(dataset_dir)
manifest = manifest_str(str(rel_data_path), audio_dur, transcript)
mf.write(manifest)
if verbose:
print(f"writing '{transcript}' of duration {audio_dur}")
num_datapoints += 1
return num_datapoints
def ui_data_generator(dataset_dir, asr_data_source, verbose=False):
(dataset_dir / Path("wavs")).mkdir(parents=True, exist_ok=True)
(dataset_dir / Path("wav_plots")).mkdir(parents=True, exist_ok=True)
def data_fn(
transcript,
audio_dur,
wav_data,
caller_name,
aud_seg,
fname,
audio_file,
num_datapoints,
rel_data_path,
):
png_path = Path(fname).with_suffix(".png")
rel_plot_path = Path("wav_plots") / png_path
wav_plot_path = dataset_dir / rel_plot_path
if not wav_plot_path.exists():
plot_seg(wav_plot_path.absolute(), audio_file)
return {
"audio_path": str(rel_data_path),
"duration": round(audio_dur, 1),
"text": transcript,
"real_idx": num_datapoints,
"caller": caller_name,
"utterance_id": fname,
"plot_path": str(rel_plot_path),
}
num_datapoints = 0
data_funcs = []
for transcript, audio_dur, wav_data, caller_name, aud_seg in asr_data_source:
fname = str(uuid4()) + "_" + slugify(transcript, max_length=8)
audio_file = (
dataset_dir / Path("wavs") / Path(fname).with_suffix(".wav")
).absolute()
audio_file.write_bytes(wav_data)
# audio_path = str(audio_file)
rel_data_path = audio_file.relative_to(dataset_dir.absolute())
data_funcs.append(
partial(
data_fn,
transcript,
audio_dur,
wav_data,
caller_name,
aud_seg,
fname,
audio_file,
num_datapoints,
rel_data_path,
)
)
num_datapoints += 1
ui_data = parallel_apply(lambda x: x(), data_funcs)
return ui_data, num_datapoints
def ui_dump_manifest_writer(dataset_dir, asr_data_source, verbose=False):
dump_data, num_datapoints = ui_data_generator(
dataset_dir, asr_data_source, verbose=verbose
)
asr_manifest = dataset_dir / Path("manifest.json")
with asr_manifest.open("w") as mf:
print(f"writing manifest to {asr_manifest}")
for d in dump_data:
rel_data_path = d["audio_path"]
audio_dur = d["duration"]
transcript = d["text"]
manifest = manifest_str(str(rel_data_path), audio_dur, transcript)
mf.write(manifest)
ui_dump_file = dataset_dir / Path("ui_dump.json")
ExtendedPath(ui_dump_file).write_json({"data": dump_data})
return num_datapoints
def asr_manifest_reader(data_manifest_path: Path):
print(f"reading manifest from {data_manifest_path}")
with data_manifest_path.open("r") as pf:
data_jsonl = pf.readlines()
data_data = [json.loads(v) for v in data_jsonl]
for p in data_data:
p["audio_path"] = data_manifest_path.parent / Path(p["audio_filepath"])
p["text"] = p["text"].strip()
yield p
def asr_manifest_writer(asr_manifest_path: Path, manifest_str_source):
with asr_manifest_path.open("w") as mf:
print(f"opening {asr_manifest_path} for writing manifest")
for mani_dict in manifest_str_source:
manifest = manifest_str(
mani_dict["audio_filepath"], mani_dict["duration"], mani_dict["text"]
)
mf.write(manifest)
def asr_test_writer(out_file_path: Path, source):
def dd_str(dd, idx):
path = dd["audio_filepath"]
# dur = dd["duration"]
# return f"SAY {idx}\nPAUSE 3\nPLAY {path}\nPAUSE 3\n\n"
return f"PAUSE 2\nPLAY {path}\nPAUSE 60\n\n"
res_file = out_file_path.with_suffix(".result.json")
with out_file_path.open("w") as of:
print(f"opening {out_file_path} for writing test")
results = []
idx = 0
for ui_dd in source:
results.append(ui_dd)
out_str = dd_str(ui_dd, idx)
of.write(out_str)
idx += 1
of.write("DO_HANGUP\n")
ExtendedPath(res_file).write_json(results)
def batch(iterable, n=1):
ls = len(iterable)
return [iterable[ndx : min(ndx + n, ls)] for ndx in range(0, ls, n)]
class ExtendedPath(type(Path())):
"""docstring for ExtendedPath."""
def read_json(self):
print(f"reading json from {self}")
with self.open("r") as jf:
return json.load(jf)
def read_yaml(self):
yaml = YAML(typ="safe", pure=True)
print(f"reading yaml from {self}")
with self.open("r") as yf:
return yaml.load(yf)
def read_jsonl(self):
print(f"reading jsonl from {self}")
with self.open("r") as jf:
for l in jf.readlines():
yield json.loads(l)
def write_json(self, data):
print(f"writing json to {self}")
self.parent.mkdir(parents=True, exist_ok=True)
with self.open("w") as jf:
json.dump(data, jf, indent=2)
def write_yaml(self, data):
yaml = YAML()
print(f"writing yaml to {self}")
with self.open("w") as yf:
yaml.dump(data, yf)
def write_jsonl(self, data):
print(f"writing jsonl to {self}")
self.parent.mkdir(parents=True, exist_ok=True)
with self.open("w") as jf:
for d in data:
jf.write(json.dumps(d) + "\n")
def get_mongo_coll(uri):
ud = pymongo.uri_parser.parse_uri(uri)
conn = pymongo.MongoClient(uri)
return conn[ud["database"]][ud["collection"]]
def get_mongo_conn(host="", port=27017, db="db", col="collection"):
mongo_host = host if host else os.environ.get("MONGO_HOST", "localhost")
mongo_uri = f"mongodb://{mongo_host}:{port}/"
return pymongo.MongoClient(mongo_uri)[db][col]
def strip_silence(sound):
from pydub.silence import detect_leading_silence
start_trim = detect_leading_silence(sound)
end_trim = detect_leading_silence(sound.reverse())
duration = len(sound)
return sound[start_trim : duration - end_trim]
def plot_seg(wav_plot_path, audio_path):
fig = plt.Figure()
ax = fig.add_subplot()
(y, sr) = librosa.load(str(audio_path))
audio_display.waveplot(y=y, sr=sr, ax=ax)
with wav_plot_path.open("wb") as wav_plot_f:
fig.set_tight_layout(True)
fig.savefig(wav_plot_f, format="png", dpi=50)
def parallel_apply(fn, iterable, workers=8, pool="thread"):
if pool == "thread":
with ThreadPoolExecutor(max_workers=workers) as exe:
print(f"parallelly applying {fn}")
return [
res
for res in tqdm(
exe.map(fn, iterable), position=0, leave=True, total=len(iterable)
)
]
elif pool == "process":
with ProcessPoolExecutor(max_workers=workers) as exe:
print(f"parallelly applying {fn}")
return [
res
for res in tqdm(
exe.map(fn, iterable), position=0, leave=True, total=len(iterable)
)
]
else:
raise Exception(f"unsupported pool type - {pool}")
def generate_filter_map(src_dataset_path, dest_dataset_path, data_file):
min_nums = 3
max_duration = 1 * 60 * 60
skip_duration = 1 * 60 * 60
def filtered_max_dur():
wav_duration = 0
for s in ExtendedPath(data_file).read_jsonl():
nums = re.sub(" ", "", s["text"])
if len(nums) >= min_nums:
wav_duration += s["duration"]
shutil.copy(
src_dataset_path / Path(s["audio_filepath"]),
dest_dataset_path / Path(s["audio_filepath"]),
)
yield s
if wav_duration > max_duration:
break
typer.echo(f"filtered only {duration_str(wav_duration)} of audio")
def filtered_skip_dur():
wav_duration = 0
for s in ExtendedPath(data_file).read_jsonl():
nums = re.sub(" ", "", s["text"])
if len(nums) >= min_nums:
wav_duration += s["duration"]
if wav_duration <= skip_duration:
continue
elif len(nums) >= min_nums:
yield s
shutil.copy(
src_dataset_path / Path(s["audio_filepath"]),
dest_dataset_path / Path(s["audio_filepath"]),
)
typer.echo(f"skipped {duration_str(skip_duration)} of audio")
def filtered_blanks():
blank_count = 0
for s in ExtendedPath(data_file).read_jsonl():
nums = re.sub(" ", "", s["text"])
if nums != "":
blank_count += 1
shutil.copy(
src_dataset_path / Path(s["audio_filepath"]),
dest_dataset_path / Path(s["audio_filepath"]),
)
yield s
typer.echo(f"filtered {blank_count} blank samples")
def filtered_transform_digits():
count = 0
for s in ExtendedPath(data_file).read_jsonl():
count += 1
digit_text = replace_digit_symbol(s["text"])
only_digits = discard_except_digits(digit_text)
char_text = digits_to_chars(only_digits)
shutil.copy(
src_dataset_path / Path(s["audio_filepath"]),
dest_dataset_path / Path(s["audio_filepath"]),
)
s["text"] = char_text
yield s
typer.echo(f"transformed {count} samples")
def filtered_extract_chars():
count = 0
for s in ExtendedPath(data_file).read_jsonl():
count += 1
no_digits = digits_to_chars(s["text"]).upper()
only_chars = re.sub("[^A-Z'\b]", " ", no_digits)
filter_text = replace_redundant_spaces_with(only_chars, " ").strip()
shutil.copy(
src_dataset_path / Path(s["audio_filepath"]),
dest_dataset_path / Path(s["audio_filepath"]),
)
s["text"] = filter_text
yield s
typer.echo(f"transformed {count} samples")
def filtered_resample():
count = 0
for s in ExtendedPath(data_file).read_jsonl():
count += 1
src_aud = pydub.AudioSegment.from_file(
src_dataset_path / Path(s["audio_filepath"])
)
dst_aud = src_aud.set_channels(1).set_sample_width(1).set_frame_rate(24000)
dst_aud.export(dest_dataset_path / Path(s["audio_filepath"]), format="wav")
yield s
typer.echo(f"transformed {count} samples")
filter_kind_map = {
"max_dur_1hr_min3num": filtered_max_dur,
"skip_dur_1hr_min3num": filtered_skip_dur,
"blanks": filtered_blanks,
"transform_digits": filtered_transform_digits,
"extract_chars": filtered_extract_chars,
"resample_ulaw24kmono": filtered_resample,
}
return filter_kind_map

117
plume/utils/align.py Normal file
View File

@@ -0,0 +1,117 @@
from pathlib import Path
from .tts import GoogleTTS
# from IPython import display
import requests
import io
import typer
from plume.utils import lazy_module
display = lazy_module('IPython.display')
pydub = lazy_module('pydub')
app = typer.Typer()
# Start gentle with following command
# docker run --rm -d --name gentle_service -p 8765:8765/tcp lowerquality/gentle
def gentle_aligner(service_uri, wav_data, utter_text):
# service_uri= "http://52.41.161.36:8765/transcriptions"
wav_f = io.BytesIO(wav_data)
wav_seg = pydub.AudioSegment.from_file(wav_f)
mp3_f = io.BytesIO()
wav_seg.export(mp3_f, format="mp3")
mp3_f.seek(0)
params = (("async", "false"),)
files = {
"audio": ("audio.mp3", mp3_f),
"transcript": ("words.txt", io.BytesIO(utter_text.encode("utf-8"))),
}
response = requests.post(service_uri, params=params, files=files)
print(f"Time duration of audio {wav_seg.duration_seconds}")
print(f"Time taken to align: {response.elapsed}s")
return wav_seg, response.json()
def gentle_align_iter(service_uri, wav_data, utter_text):
wav_seg, response = gentle_aligner(service_uri, wav_data, utter_text)
for span in response:
word_seg = wav_seg[int(span["start"] * 1000) : int(span["end"] * 1000)]
word = span["word"]
yield (word, word_seg)
def tts_jupyter():
google_voices = GoogleTTS.voice_list()
gtts = GoogleTTS()
# google_voices[4]
us_voice = [v for v in google_voices if v["language"] == "en-US"][0]
utter_text = (
"I would like to align the audio segments based on word level timestamps"
)
wav_data = gtts.text_to_speech(text=utter_text, params=us_voice)
for word, seg in gentle_align_iter(wav_data, utter_text):
print(word)
display.display(seg)
@app.command()
def cut(audio_path: Path, transcript_path: Path, out_dir: Path = "/tmp"):
from . import ExtendedPath
import datetime
import re
aud_seg = pydub.AudioSegment.from_file(audio_path)
aud_seg[: 15 * 60 * 1000].export(out_dir / Path("audio.mp3"), format="mp3")
tscript_json = ExtendedPath(transcript_path).read_json()
def time_to_msecs(time_str):
return (
datetime.datetime.strptime(time_str, "%H:%M:%S,%f")
- datetime.datetime(1900, 1, 1)
).total_seconds() * 1000
tscript_words = []
broken = False
for m in tscript_json["monologues"]:
# tscript_words.append("|")
for e in m["elements"]:
if e["type"] == "text":
text = e["value"]
text = re.sub(r"\[.*\]", "", text)
text = re.sub(r"\(.*\)", "", text)
tscript_words.append(text)
if "timestamp" in e and time_to_msecs(e["timestamp"]) >= 15 * 60 * 1000:
broken = True
break
if broken:
break
(out_dir / Path("words.txt")).write_text("".join(tscript_words))
@app.command()
def gentle_preview(
audio_path: Path,
transcript_path: Path,
service_uri="http://101.53.142.218:8765/transcriptions",
gent_preview_dir="../gentle_preview",
):
from . import ExtendedPath
ab = audio_path.read_bytes()
tt = transcript_path.read_text()
audio, alignment = gentle_aligner(service_uri, ab, tt)
audio.export(gent_preview_dir / Path("a.wav"), format="wav")
alignment["status"] = "OK"
ExtendedPath(gent_preview_dir / Path("status.json")).write_json(alignment)
def main():
app()
if __name__ == "__main__":
main()

28
plume/utils/audio.py Normal file
View File

@@ -0,0 +1,28 @@
from scipy.signal import lfilter, butter
from scipy.io.wavfile import read, write
from numpy import array, int16
import sys
def butter_params(low_freq, high_freq, fs, order=5):
nyq = 0.5 * fs
low = low_freq / nyq
high = high_freq / nyq
b, a = butter(order, [low, high], btype="band")
return b, a
def butter_bandpass_filter(data, low_freq, high_freq, fs, order=5):
b, a = butter_params(low_freq, high_freq, fs, order=order)
y = lfilter(b, a, data)
return y
if __name__ == "__main__":
fs, audio = read(sys.argv[1])
import pdb; pdb.set_trace()
low_freq = 300.0
high_freq = 4000.0
filtered_signal = butter_bandpass_filter(audio, low_freq, high_freq, fs, order=6)
fname = sys.argv[1].split(".wav")[0] + "_moded.wav"
write(fname, fs, array(filtered_signal, dtype=int16))

737
plume/utils/lazy_import.py Normal file
View File

@@ -0,0 +1,737 @@
# -*- Mode: python; tab-width: 4; indent-tabs-mode:nil; coding:utf-8 -*-
# vim: tabstop=4 expandtab shiftwidth=4 softtabstop=4
#
# lazy_import --- https://github.com/mnmelo/lazy_import
# Copyright (C) 2017-2018 Manuel Nuno Melo
#
# This file is part of lazy_import.
#
# lazy_import is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# lazy_import is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with lazy_import. If not, see <http://www.gnu.org/licenses/>.
#
# lazy_import was based on code from the importing module from the PEAK
# package (see <http://peak.telecommunity.com/DevCenter/Importing>). The PEAK
# package is released under the following license, reproduced here:
#
# Copyright (C) 1996-2004 by Phillip J. Eby and Tyler C. Sarna.
# All rights reserved. This software may be used under the same terms
# as Zope or Python. THERE ARE ABSOLUTELY NO WARRANTIES OF ANY KIND.
# Code quality varies between modules, from "beta" to "experimental
# pre-alpha". :)
#
# Code pertaining to lazy loading from PEAK importing was included in
# lazy_import, modified in a number of ways. These are detailed in the
# CHANGELOG file of lazy_import. Changes mainly involved Python 3
# compatibility, extension to allow customizable behavior, and added
# functionality (lazy importing of callable objects).
#
"""
Lazy module loading
===================
Functions and classes for lazy module loading that also delay import errors.
Heavily borrowed from the `importing`_ module.
.. _`importing`: http://peak.telecommunity.com/DevCenter/Importing
Files and directories
---------------------
.. autofunction:: module
.. autofunction:: callable
"""
__all__ = [
"lazy_module",
"lazy_callable",
"lazy_function",
"lazy_class",
"LazyModule",
"LazyCallable",
"module_basename",
"_MSG",
"_MSG_CALLABLE",
]
from types import ModuleType
import sys
try:
from importlib._bootstrap import _ImportLockContext
except ImportError:
# Python 2 doesn't have the context manager. Roll it ourselves (copied from
# Python 3's importlib/_bootstrap.py)
import imp
class _ImportLockContext:
"""Context manager for the import lock."""
def __enter__(self):
imp.acquire_lock()
def __exit__(self, exc_type, exc_value, exc_traceback):
imp.release_lock()
# Adding a __spec__ doesn't really help. I'll leave the code here in case
# future python implementations start relying on it.
# try:
# from importlib.machinery import ModuleSpec
# except ImportError:
# ModuleSpec = None
import six
from six import raise_from
from six.moves import reload_module
# It is sometime useful to have access to the version number of a library.
# This is usually done through the __version__ special attribute.
# To make sure the version number is consistent between setup.py and the
# library, we read the version number from the file called VERSION that stays
# in the module directory.
import os
# VERSION_FILE = os.path.join(os.path.dirname(__file__), "VERSION")
# with open(VERSION_FILE) as infile:
# __version__ = infile.read().strip()
# Logging
import logging
# adding a TRACE level for stack debugging
_LAZY_TRACE = 1
logging.addLevelName(1, "LAZY_TRACE")
logging.basicConfig(level=logging.WARNING)
# Logs a formatted stack (takes no message or args/kwargs)
def _lazy_trace(self):
if self.isEnabledFor(_LAZY_TRACE):
import traceback
self._log(_LAZY_TRACE, " ### STACK TRACE ###", ())
for line in traceback.format_stack(sys._getframe(2)):
for subline in line.split("\n"):
self._log(_LAZY_TRACE, subline.rstrip(), ())
logging.Logger.lazy_trace = _lazy_trace
logger = logging.getLogger(__name__)
################################
# Module/function registration #
################################
#### Lazy classes ####
class LazyModule(ModuleType):
"""Class for lazily-loaded modules that triggers proper loading on access.
Instantiation should be made from a subclass of :class:`LazyModule`, with
one subclass per instantiated module. Regular attribute set/access can then
be recovered by setting the subclass's :meth:`__getattribute__` and
:meth:`__setattribute__` to those of :class:`types.ModuleType`.
"""
# peak.util.imports sets __slots__ to (), but it seems pointless because
# the base ModuleType doesn't itself set __slots__.
def __getattribute__(self, attr):
logger.debug(
"Getting attr {} of LazyModule instance of {}".format(
attr, super(LazyModule, self).__getattribute__("__name__")
)
)
logger.lazy_trace()
# IPython tries to be too clever and constantly inspects, asking for
# modules' attrs, which causes premature module loading and unesthetic
# internal errors if the lazily-loaded module doesn't exist.
if (
run_from_ipython()
and (attr.startswith(("__", "_ipython")) or attr == "_repr_mimebundle_")
and module_basename(_caller_name()) in ("inspect", "IPython")
):
logger.debug(
"Ignoring request for {}, deemed from IPython's "
"inspection.".format(
super(LazyModule, self).__getattribute__("__name__"), attr
)
)
raise AttributeError
if not attr in ("__name__", "__class__", "__spec__"):
# __name__ and __class__ yield their values from the LazyModule;
# __spec__ causes an AttributeError. Maybe in the future it will be
# necessary to return an actual ModuleSpec object, but it works as
# it is without that now.
# If it's an already-loaded submodule, we return it without
# triggering a full loading
try:
return sys.modules[self.__name__ + "." + attr]
except KeyError:
pass
# Check if it's one of the lazy callables
try:
_callable = type(self)._lazy_import_callables[attr]
logger.debug("Returning lazy-callable '{}'.".format(attr))
return _callable
except (AttributeError, KeyError) as err:
logger.debug(
"Proceeding to load module {}, "
"from requested value {}".format(
super(LazyModule, self).__getattribute__("__name__"), attr
)
)
_load_module(self)
logger.debug(
"Returning value '{}'.".format(
super(LazyModule, self).__getattribute__(attr)
)
)
return super(LazyModule, self).__getattribute__(attr)
def __setattr__(self, attr, value):
logger.debug(
"Setting attr {} to value {}, in LazyModule instance "
"of {}".format(
attr, value, super(LazyModule, self).__getattribute__("__name__")
)
)
_load_module(self)
return super(LazyModule, self).__setattr__(attr, value)
class LazyCallable(object):
"""Class for lazily-loaded callables that triggers module loading on access
"""
def __init__(self, *args):
if len(args) != 2:
# Maybe the user tried to base a class off this lazy callable?
try:
logger.debug(
"Got wrong number of args when init'ing "
"LazyCallable. args is '{}'".format(args)
)
base = args[1][0]
if isinstance(base, LazyCallable) and len(args) == 3:
raise NotImplementedError(
"It seems you are trying to use "
"a lazy callable as a class "
"base. This is not supported."
)
except (IndexError, TypeError):
raise_from(
TypeError(
"LazyCallable takes exactly 2 arguments: "
"a module/lazy module object and the name of "
"a callable to be lazily loaded."
),
None,
)
self.module, self.cname = args
self.modclass = type(self.module)
self.callable = None
# Need to save these, since the module-loading gets rid of them
self.error_msgs = self.modclass._lazy_import_error_msgs
self.error_strings = self.modclass._lazy_import_error_strings
def __call__(self, *args, **kwargs):
# No need to go through all the reloading more than once.
if self.callable:
return self.callable(*args, **kwargs)
try:
del self.modclass._lazy_import_callables[self.cname]
except (AttributeError, KeyError):
pass
try:
self.callable = getattr(self.module, self.cname)
except AttributeError:
msg = self.error_msgs["msg_callable"]
raise_from(
AttributeError(msg.format(callable=self.cname, **self.error_strings)),
None,
)
except ImportError as err:
# Import failed. We reset the dict and re-raise the ImportError.
try:
self.modclass._lazy_import_callables[self.cname] = self
except AttributeError:
self.modclass._lazy_import_callables = {self.cname: self}
raise_from(err, None)
else:
return self.callable(*args, **kwargs)
### Functions ###
def lazy_module(modname, error_strings=None, lazy_mod_class=LazyModule, level="leaf"):
"""Function allowing lazy importing of a module into the namespace.
A lazy module object is created, registered in `sys.modules`, and
returned. This is a hollow module; actual loading, and `ImportErrors` if
not found, are delayed until an attempt is made to access attributes of the
lazy module.
A handy application is to use :func:`lazy_module` early in your own code
(say, in `__init__.py`) to register all modulenames you want to be lazy.
Because of registration in `sys.modules` later invocations of
`import modulename` will also return the lazy object. This means that after
initial registration the rest of your code can use regular pyhon import
statements and retain the lazyness of the modules.
Parameters
----------
modname : str
The module to import.
error_strings : dict, optional
A dictionary of strings to use when module-loading fails. Key 'msg'
sets the message to use (defaults to :attr:`lazy_import._MSG`). The
message is formatted using the remaining dictionary keys. The default
message informs the user of which module is missing (key 'module'),
what code loaded the module as lazy (key 'caller'), and which package
should be installed to solve the dependency (key 'install_name').
None of the keys is mandatory and all are given smart names by default.
lazy_mod_class: type, optional
Which class to use when instantiating the lazy module, to allow
deep customization. The default is :class:`LazyModule` and custom
alternatives **must** be a subclass thereof.
level : str, optional
Which submodule reference to return. Either a reference to the 'leaf'
module (the default) or to the 'base' module. This is useful if you'll
be using the module functionality in the same place you're calling
:func:`lazy_module` from, since then you don't need to run `import`
again. Setting *level* does not affect which names/modules get
registered in `sys.modules`.
For *level* set to 'base' and *modulename* 'aaa.bbb.ccc'::
aaa = lazy_import.lazy_module("aaa.bbb.ccc", level='base')
# 'aaa' becomes defined in the current namespace, with
# (sub)attributes 'aaa.bbb' and 'aaa.bbb.ccc'.
# It's the lazy equivalent to:
import aaa.bbb.ccc
For *level* set to 'leaf'::
ccc = lazy_import.lazy_module("aaa.bbb.ccc", level='leaf')
# Only 'ccc' becomes set in the current namespace.
# Lazy equivalent to:
from aaa.bbb import ccc
Returns
-------
module
The module specified by *modname*, or its base, depending on *level*.
The module isn't immediately imported. Instead, an instance of
*lazy_mod_class* is returned. Upon access to any of its attributes, the
module is finally loaded.
Examples
--------
>>> import lazy_import, sys
>>> np = lazy_import.lazy_module("numpy")
>>> np
Lazily-loaded module numpy
>>> np is sys.modules['numpy']
True
>>> np.pi # This causes the full loading of the module ...
3.141592653589793
>>> np # ... and the module is changed in place.
<module 'numpy' from '/usr/local/lib/python/site-packages/numpy/__init__.py'>
>>> import lazy_import, sys
>>> # The following succeeds even when asking for a module that's not available
>>> missing = lazy_import.lazy_module("missing_module")
>>> missing
Lazily-loaded module missing_module
>>> missing is sys.modules['missing_module']
True
>>> missing.some_attr # This causes the full loading of the module, which now fails.
ImportError: __main__ attempted to use a functionality that requires module missing_module, but it couldn't be loaded. Please install missing_module and retry.
See Also
--------
:func:`lazy_callable`
:class:`LazyModule`
"""
if error_strings is None:
error_strings = {}
_set_default_errornames(modname, error_strings)
mod = _lazy_module(modname, error_strings, lazy_mod_class)
if level == "base":
return sys.modules[module_basename(modname)]
elif level == "leaf":
return mod
else:
raise ValueError("Parameter 'level' must be one of ('base', 'leaf')")
def _lazy_module(modname, error_strings, lazy_mod_class):
with _ImportLockContext():
fullmodname = modname
fullsubmodname = None
# ensure parent module/package is in sys.modules
# and parent.modname=module, as soon as the parent is imported
while modname:
try:
mod = sys.modules[modname]
# We reached a (base) module that's already loaded. Let's stop
# the cycle. Can't use 'break' because we still want to go
# through the fullsubmodname check below.
modname = ""
except KeyError:
err_s = error_strings.copy()
err_s.setdefault("module", modname)
class _LazyModule(lazy_mod_class):
_lazy_import_error_msgs = {"msg": err_s.pop("msg")}
try:
_lazy_import_error_msgs["msg_callable"] = err_s.pop(
"msg_callable"
)
except KeyError:
pass
_lazy_import_error_strings = err_s
_lazy_import_callables = {}
_lazy_import_submodules = {}
def __repr__(self):
return "Lazily-loaded module {}".format(self.__name__)
# A bit of cosmetic, to make AttributeErrors read more natural
_LazyModule.__name__ = "module"
# Actual module instantiation
mod = sys.modules[modname] = _LazyModule(modname)
# No need for __spec__. Maybe in the future.
# if ModuleSpec:
# ModuleType.__setattr__(mod, '__spec__',
# ModuleSpec(modname, None))
if fullsubmodname:
submod = sys.modules[fullsubmodname]
ModuleType.__setattr__(mod, submodname, submod)
_LazyModule._lazy_import_submodules[submodname] = submod
fullsubmodname = modname
modname, _, submodname = modname.rpartition(".")
return sys.modules[fullmodname]
def lazy_callable(modname, *names, **kwargs):
"""Performs lazy importing of one or more callables.
:func:`lazy_callable` creates functions that are thin wrappers that pass
any and all arguments straight to the target module's callables. These can
be functions or classes. The full loading of that module is only actually
triggered when the returned lazy function itself is called. This lazy
import of the target module uses the same mechanism as
:func:`lazy_module`.
If, however, the target module has already been fully imported prior
to invocation of :func:`lazy_callable`, then the target callables
themselves are returned and no lazy imports are made.
:func:`lazy_function` and :func:`lazy_function` are aliases of
:func:`lazy_callable`.
Parameters
----------
modname : str
The base module from where to import the callable(s) in *names*,
or a full 'module_name.callable_name' string.
names : str (optional)
The callable name(s) to import from the module specified by *modname*.
If left empty, *modname* is assumed to also include the callable name
to import.
error_strings : dict, optional
A dictionary of strings to use when reporting loading errors (either a
missing module, or a missing callable name in the loaded module).
*error_string* follows the same usage as described under
:func:`lazy_module`, with the exceptions that 1) a further key,
'msg_callable', can be supplied to be used as the error when a module
is successfully loaded but the target callable can't be found therein
(defaulting to :attr:`lazy_import._MSG_CALLABLE`); 2) a key 'callable'
is always added with the callable name being loaded.
lazy_mod_class : type, optional
See definition under :func:`lazy_module`.
lazy_call_class : type, optional
Analogously to *lazy_mod_class*, allows setting a custom class to
handle lazy callables, other than the default :class:`LazyCallable`.
Returns
-------
wrapper function or tuple of wrapper functions
If *names* is passed, returns a tuple of wrapper functions, one for
each element in *names*.
If only *modname* is passed it is assumed to be a full
'module_name.callable_name' string, in which case the wrapper for the
imported callable is returned directly, and not in a tuple.
Notes
-----
Unlike :func:`lazy_module`, which returns a lazy module that eventually
mutates into the fully-functional version, :func:`lazy_callable` only
returns thin wrappers that never change. This means that the returned
wrapper object never truly becomes the one under the module's namespace,
even after successful loading of the module in *modname*. This is fine for
most practical use cases, but may break code that relies on the usage of
the returned objects oter than calling them. One such example is the lazy
import of a class: it's fine to use the returned wrapper to instantiate an
object, but it can't be used, for instance, to subclass from.
Examples
--------
>>> import lazy_import, sys
>>> fn = lazy_import.lazy_callable("numpy.arange")
>>> sys.modules['numpy']
Lazily-loaded module numpy
>>> fn(10)
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> sys.modules['numpy']
<module 'numpy' from '/usr/local/lib/python3.5/site-packages/numpy/__init__.py'>
>>> import lazy_import, sys
>>> cl = lazy_import.lazy_callable("numpy.ndarray") # a class
>>> obj = cl([1, 2]) # This works OK (and also triggers the loading of numpy)
>>> class MySubclass(cl): # This fails because cls is just a wrapper,
>>> pass # not an actual class.
See Also
--------
:func:`lazy_module`
:class:`LazyCallable`
:class:`LazyModule`
"""
if not names:
modname, _, name = modname.rpartition(".")
lazy_mod_class = _setdef(kwargs, "lazy_mod_class", LazyModule)
lazy_call_class = _setdef(kwargs, "lazy_call_class", LazyCallable)
error_strings = _setdef(kwargs, "error_strings", {})
_set_default_errornames(modname, error_strings, call=True)
if not names:
# We allow passing a single string as 'modname.callable_name',
# in which case the wrapper is returned directly and not as a list.
return _lazy_callable(
modname, name, error_strings.copy(), lazy_mod_class, lazy_call_class
)
return tuple(
_lazy_callable(
modname, cname, error_strings.copy(), lazy_mod_class, lazy_call_class
)
for cname in names
)
lazy_function = lazy_class = lazy_callable
def _lazy_callable(modname, cname, error_strings, lazy_mod_class, lazy_call_class):
# We could do most of this in the LazyCallable __init__, but here we can
# pre-check whether to actually be lazy or not.
module = _lazy_module(modname, error_strings, lazy_mod_class)
modclass = type(module)
if issubclass(modclass, LazyModule) and hasattr(modclass, "_lazy_import_callables"):
modclass._lazy_import_callables.setdefault(
cname, lazy_call_class(module, cname)
)
return getattr(module, cname)
#######################
# Real module loading #
#######################
def _load_module(module):
"""Ensures that a module, and its parents, are properly loaded
"""
modclass = type(module)
# We only take care of our own LazyModule instances
if not issubclass(modclass, LazyModule):
raise TypeError("Passed module is not a LazyModule instance.")
with _ImportLockContext():
parent, _, modname = module.__name__.rpartition(".")
logger.debug("loading module {}".format(modname))
# We first identify whether this is a loadable LazyModule, then we
# strip as much of lazy_import behavior as possible (keeping it cached,
# in case loading fails and we need to reset the lazy state).
if not hasattr(modclass, "_lazy_import_error_msgs"):
# Alreay loaded (no _lazy_import_error_msgs attr). Not reloading.
return
# First, ensure the parent is loaded (using recursion; *very* unlikely
# we'll ever hit a stack limit in this case).
modclass._LOADING = True
try:
if parent:
logger.debug("first loading parent module {}".format(parent))
setattr(sys.modules[parent], modname, module)
if not hasattr(modclass, "_LOADING"):
logger.debug("Module {} already loaded by the parent".format(modname))
# We've been loaded by the parent. Let's bail.
return
cached_data = _clean_lazymodule(module)
try:
# Get Python to do the real import!
reload_module(module)
except:
# Loading failed. We reset our lazy state.
logger.debug("Failed to load module {}. Resetting...".format(modname))
_reset_lazymodule(module, cached_data)
raise
else:
# Successful load
logger.debug("Successfully loaded module {}".format(modname))
delattr(modclass, "_LOADING")
_reset_lazy_submod_refs(module)
except (AttributeError, ImportError) as err:
logger.debug(
"Failed to load {}.\n{}: {}".format(
modname, err.__class__.__name__, err
)
)
logger.lazy_trace()
# Under Python 3 reloading our dummy LazyModule instances causes an
# AttributeError if the module can't be found. Would be preferrable
# if we could always rely on an ImportError. As it is we vet the
# AttributeError as thoroughly as possible.
if (six.PY3 and isinstance(err, AttributeError)) and not err.args[
0
] == "'NoneType' object has no attribute 'name'":
# Not the AttributeError we were looking for.
raise
msg = modclass._lazy_import_error_msgs["msg"]
raise_from(
ImportError(msg.format(**modclass._lazy_import_error_strings)), None
)
##############################
# Helper functions/constants #
##############################
_MSG = (
"{caller} attempted to use a functionality that requires module "
"{module}, but it couldn't be loaded. Please install {install_name} "
"and retry."
)
_MSG_CALLABLE = (
"{caller} attempted to use a functionality that requires "
"{callable}, of module {module}, but it couldn't be found in that "
"module. Please install a version of {install_name} that has "
"{module}.{callable} and retry."
)
_CLS_ATTRS = (
"_lazy_import_error_strings",
"_lazy_import_error_msgs",
"_lazy_import_callables",
"_lazy_import_submodules",
"__repr__",
)
_DELETION_DICT = ("_lazy_import_submodules",)
def _setdef(argdict, name, defaultvalue):
"""Like dict.setdefault but sets the default value also if None is present.
"""
if not name in argdict or argdict[name] is None:
argdict[name] = defaultvalue
return argdict[name]
def module_basename(modname):
return modname.partition(".")[0]
def _set_default_errornames(modname, error_strings, call=False):
# We don't set the modulename default here because it will change for
# parents of lazily imported submodules.
error_strings.setdefault("caller", _caller_name(3, default="Python"))
error_strings.setdefault("install_name", module_basename(modname))
error_strings.setdefault("msg", _MSG)
if call:
error_strings.setdefault("msg_callable", _MSG_CALLABLE)
def _caller_name(depth=2, default=""):
"""Returns the name of the calling namespace.
"""
# the presence of sys._getframe might be implementation-dependent.
# It isn't that serious if we can't get the caller's name.
try:
return sys._getframe(depth).f_globals["__name__"]
except AttributeError:
return default
def _clean_lazymodule(module):
"""Removes all lazy behavior from a module's class, for loading.
Also removes all module attributes listed under the module's class deletion
dictionaries. Deletion dictionaries are class attributes with names
specified in `_DELETION_DICT`.
Parameters
----------
module: LazyModule
Returns
-------
dict
A dictionary of deleted class attributes, that can be used to reset the
lazy state using :func:`_reset_lazymodule`.
"""
modclass = type(module)
_clean_lazy_submod_refs(module)
modclass.__getattribute__ = ModuleType.__getattribute__
modclass.__setattr__ = ModuleType.__setattr__
cls_attrs = {}
for cls_attr in _CLS_ATTRS:
try:
cls_attrs[cls_attr] = getattr(modclass, cls_attr)
delattr(modclass, cls_attr)
except AttributeError:
pass
return cls_attrs
def _clean_lazy_submod_refs(module):
modclass = type(module)
for deldict in _DELETION_DICT:
try:
delnames = getattr(modclass, deldict)
except AttributeError:
continue
for delname in delnames:
try:
super(LazyModule, module).__delattr__(delname)
except AttributeError:
# Maybe raise a warning?
pass
def _reset_lazymodule(module, cls_attrs):
"""Resets a module's lazy state from cached data.
"""
modclass = type(module)
del modclass.__getattribute__
del modclass.__setattr__
try:
del modclass._LOADING
except AttributeError:
pass
for cls_attr in _CLS_ATTRS:
try:
setattr(modclass, cls_attr, cls_attrs[cls_attr])
except KeyError:
pass
_reset_lazy_submod_refs(module)
def _reset_lazy_submod_refs(module):
modclass = type(module)
for deldict in _DELETION_DICT:
try:
resetnames = getattr(modclass, deldict)
except AttributeError:
continue
for name, submod in resetnames.items():
super(LazyModule, module).__setattr__(name, submod)
def run_from_ipython():
# Taken from https://stackoverflow.com/questions/5376837
try:
__IPYTHON__
return True
except NameError:
return False

46
plume/utils/lazy_loader.py Normal file
View File

@@ -0,0 +1,46 @@
# Code copied from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/util/lazy_loader.py
"""A LazyLoader class."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import importlib
import types
class LazyLoader(types.ModuleType):
"""Lazily import a module, mainly to avoid pulling in large dependencies.
`contrib`, and `ffmpeg` are examples of modules that are large and not always
needed, and this allows them to only be loaded when they are used.
"""
# The lint error here is incorrect.
def __init__(
self, local_name, parent_module_globals, name
): # pylint: disable=super-on-old-class
self._local_name = local_name
self._parent_module_globals = parent_module_globals
super(LazyLoader, self).__init__(name)
def _load(self):
# Import the target module and insert it into the parent's namespace
module = importlib.import_module(self.__name__)
self._parent_module_globals[self._local_name] = module
# Update this object's dict so that if someone keeps a reference to the
# LazyLoader, lookups are efficient (__getattr__ is only called on lookups
# that fail).
self.__dict__.update(module.__dict__)
return module
def __getattr__(self, item):
module = self._load()
return getattr(module, item)
def __dir__(self):
module = self._load()
return dir(module)

31
plume/utils/serve.py Normal file
View File

@@ -0,0 +1,31 @@
from plume.utils import lazy_module
import typer
rpyc = lazy_module('rpyc')
app = typer.Typer()
class ASRService(rpyc.Service):
def __init__(self, asr_recognizer):
self.asr = asr_recognizer
def on_connect(self, conn):
# code that runs when a connection is created
# (to init the service, if needed)
pass
def on_disconnect(self, conn):
# code that runs after the connection has already closed
# (to finalize the service, if needed)
pass
def exposed_transcribe(self, utterance: bytes): # this is an exposed method
speech_audio = self.asr.transcribe(utterance)
return speech_audio
def exposed_transcribe_cb(
self, utterance: bytes, respond
): # this is an exposed method
speech_audio = self.asr.transcribe(utterance)
respond(speech_audio)

184
plume/utils/transcribe.py Normal file
View File

@@ -0,0 +1,184 @@
import os
import logging
from io import BytesIO
from pathlib import Path
from functools import lru_cache
import typer
# import rpyc
# from tqdm import tqdm
# from pydub import AudioSegment
# from pydub.silence import split_on_silence
from plume.utils import lazy_module, lazy_callable
rpyc = lazy_module('rpyc')
AudioSegment = lazy_callable('pydub.AudioSegment')
split_on_silence = lazy_callable('pydub.silence.split_on_silence')
app = typer.Typer()
logging.basicConfig(
level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s"
)
logger = logging.getLogger(__name__)
ASR_RPYC_HOST = os.environ.get("JASR_RPYC_HOST", "localhost")
ASR_RPYC_PORT = int(os.environ.get("ASR_RPYC_PORT", "8044"))
TRITON_ASR_MODEL = os.environ.get("TRITON_ASR_MODEL", "slu_wav2vec2")
TRITON_GRPC_ASR_HOST = os.environ.get("TRITON_GRPC_ASR_HOST", "localhost")
TRITON_GRPC_ASR_PORT = int(os.environ.get("TRITON_GRPC_ASR_PORT", "8001"))
@lru_cache()
def transcribe_rpyc_gen(asr_host=ASR_RPYC_HOST, asr_port=ASR_RPYC_PORT):
logger.info(f"connecting to asr server at {asr_host}:{asr_port}")
try:
asr = rpyc.connect(asr_host, asr_port).root
logger.info(f"connected to asr server successfully")
except ConnectionRefusedError:
raise Exception("env-var JASPER_ASR_RPYC_HOST invalid")
def audio_prep(aud_seg):
asr_seg = aud_seg.set_channels(1).set_sample_width(2).set_frame_rate(16000)
return asr_seg
return asr.transcribe, audio_prep
def triton_transcribe_grpc_gen(
asr_host=TRITON_GRPC_ASR_HOST,
asr_port=TRITON_GRPC_ASR_PORT,
asr_model=TRITON_ASR_MODEL,
method="chunked",
chunk_msec=5000,
sil_msec=500,
# overlap=False,
sep=" ",
):
from tritonclient.utils import np_to_triton_dtype
import tritonclient.grpc as grpcclient
import numpy as np
sup_meth = ["chunked", "silence", "whole"]
if method not in sup_meth:
meths = "|".join(sup_meth)
raise Exception(f"unsupported method {method}. pick one of {meths}")
client = grpcclient.InferenceServerClient(f"{asr_host}:{asr_port}")
def transcriber(aud_seg):
af = BytesIO()
aud_seg.export(af, format="wav")
input_audio_bytes = af.getvalue()
input_audio_data = np.array([input_audio_bytes])
inputs = [
grpcclient.InferInput(
"INPUT_AUDIO",
input_audio_data.shape,
np_to_triton_dtype(input_audio_data.dtype),
)
]
inputs[0].set_data_from_numpy(input_audio_data)
outputs = [grpcclient.InferRequestedOutput("OUTPUT_TEXT")]
response = client.infer(asr_model, inputs, request_id=str(1), outputs=outputs)
transcript = response.as_numpy("OUTPUT_TEXT")[0]
return transcript.decode("utf-8")
def chunked_transcriber(aud_seg):
if method == "silence":
sil_chunks = split_on_silence(
aud_seg,
min_silence_len=sil_msec,
silence_thresh=-50,
keep_silence=500,
)
chunks = [sc for c in sil_chunks for sc in c[::chunk_msec]]
else:
chunks = aud_seg[::chunk_msec]
# if overlap:
# chunks = [
# aud_seg[start, end]
# for start, end in range(0, int(aud_seg.duration_seconds * 1000, 1000))
# ]
# pass
transcript_list = []
sil_pad = AudioSegment.silent(duration=sil_msec)
for seg in chunks:
t_seg = sil_pad + seg + sil_pad
c_transcript = transcriber(t_seg)
transcript_list.append(c_transcript)
transcript = sep.join(transcript_list)
return transcript
def audio_prep(aud_seg):
asr_seg = aud_seg.set_channels(1).set_sample_width(2).set_frame_rate(16000)
return asr_seg
whole_transcriber = transcriber if method == "whole" else chunked_transcriber
return whole_transcriber, audio_prep
@app.command()
def file(audio_file: Path, write_file: bool = False, chunked=True):
from pydub import AudioSegment
aseg = AudioSegment.from_file(audio_file)
transcriber, prep = triton_transcribe_grpc_gen()
transcription = transcriber(prep(aseg))
typer.echo(transcription)
if write_file:
tscript_file_path = audio_file.with_suffix(".txt")
with open(tscript_file_path, "w") as tf:
tf.write(transcription)
@app.command()
def benchmark(audio_file: Path):
from pydub import AudioSegment
transcriber, audio_prep = transcribe_rpyc_gen()
file_seg = AudioSegment.from_file(audio_file)
aud_seg = audio_prep(file_seg)
def timeinfo():
from timeit import Timer
timer = Timer(lambda: transcriber(aud_seg))
number = 100
repeat = 10
time_taken = timer.repeat(repeat, number=number)
best = min(time_taken) * 1000 / number
print(f"{number} loops, best of {repeat}: {best:.3f} msec per loop")
timeinfo()
import time
time.sleep(5)
transcriber, audio_prep = triton_transcribe_grpc_gen()
aud_seg = audio_prep(file_seg)
def timeinfo():
from timeit import Timer
timer = Timer(lambda: transcriber(aud_seg))
number = 100
repeat = 10
time_taken = timer.repeat(repeat, number=number)
best = min(time_taken) * 1000 / number
print(f"{number} loops, best of {repeat}: {best:.3f} msec per loop")
timeinfo()
def main():
app()
if __name__ == "__main__":
main()

92
plume/utils/tts.py Normal file
View File

@@ -0,0 +1,92 @@
from logging import getLogger
from plume.utils import lazy_module
from pathlib import Path
import typer
# from google.cloud import texttospeech
texttospeech = lazy_module('google.cloud.texttospeech')
LOGGER = getLogger("googletts")
app = typer.Typer()
class GoogleTTS(object):
def __init__(self):
self.client = texttospeech.TextToSpeechClient()
def text_to_speech(self, text: str, params: dict) -> bytes:
tts_input = texttospeech.types.SynthesisInput(text=text)
voice = texttospeech.types.VoiceSelectionParams(
language_code=params["language"], name=params["name"]
)
audio_config = texttospeech.types.AudioConfig(
audio_encoding=texttospeech.enums.AudioEncoding.LINEAR16,
sample_rate_hertz=params["sample_rate"],
)
response = self.client.synthesize_speech(tts_input, voice, audio_config)
audio_content = response.audio_content
return audio_content
def ssml_to_speech(self, text: str, params: dict) -> bytes:
tts_input = texttospeech.types.SynthesisInput(ssml=text)
voice = texttospeech.types.VoiceSelectionParams(
language_code=params["language"], name=params["name"]
)
audio_config = texttospeech.types.AudioConfig(
audio_encoding=texttospeech.enums.AudioEncoding.LINEAR16,
sample_rate_hertz=params["sample_rate"],
)
response = self.client.synthesize_speech(tts_input, voice, audio_config)
audio_content = response.audio_content
return audio_content
@classmethod
def voice_list(cls):
"""Lists the available voices."""
client = cls().client
# Performs the list voices request
voices = client.list_voices()
results = []
for voice in voices.voices:
supported_eng_langs = [
lang for lang in voice.language_codes if lang[:2] == "en"
]
if len(supported_eng_langs) > 0:
lang = ",".join(supported_eng_langs)
else:
continue
ssml_gender = texttospeech.enums.SsmlVoiceGender(voice.ssml_gender)
results.append(
{
"name": voice.name,
"language": lang,
"gender": ssml_gender.name,
"engine": "wavenet" if "Wav" in voice.name else "standard",
"sample_rate": voice.natural_sample_rate_hertz,
}
)
return results
@app.command()
def generate_audio_file(text, dest_path: Path = "./tts_audio.wav", voice="en-US-Wavenet-D"):
tts = GoogleTTS()
selected_voice = [v for v in tts.voice_list() if v["name"] == voice][0]
wav_data = tts.text_to_speech(text, selected_voice)
with dest_path.open("wb") as wf:
wf.write(wav_data)
def main():
app()
if __name__ == "__main__":
main()