EngineX-Iluvatar/enginex-mr_series-sherpa-onnx
Archived
16
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Refactor hotwords，support loading hotwords from file (#296)

Browse Source
This commit is contained in:
Wei Kang
2023-09-14 19:33:17 +08:00
committed by GitHub
parent 087367d7fe
commit 47184f9db7
34 changed files with 803 additions and 300 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
python-api-examples/non_streaming_server.py
View File

@@ -326,6 +326,31 @@ def add_modified_beam_search_args(parser: argparse.ArgumentParser):
)
def add_hotwords_args(parser: argparse.ArgumentParser):
parser.add_argument(
"--hotwords-file",
type=str,
default="",
help="""
The file containing hotwords, one words/phrases per line, and for each
phrase the bpe/cjkchar are separated by a space. For example:
▁HE LL O ▁WORLD
你 好 世 界
""",
)
parser.add_argument(
"--hotwords-score",
type=float,
default=1.5,
help="""
The hotword score of each token for biasing word/phrase. Used only if
--hotwords-file is given.
""",
)
def check_args(args):
if not Path(args.tokens).is_file():
raise ValueError(f"{args.tokens} does not exist")
@@ -342,6 +367,10 @@ def check_args(args):
assert Path(args.decoder).is_file(), args.decoder
assert Path(args.joiner).is_file(), args.joiner
if args.hotwords_file != "":
assert args.decoding_method == "modified_beam_search", args.decoding_method
assert Path(args.hotwords_file).is_file(), args.hotwords_file
def get_args():
parser = argparse.ArgumentParser(
@@ -351,6 +380,7 @@ def get_args():
add_model_args(parser)
add_feature_config_args(parser)
add_decoding_args(parser)
add_hotwords_args(parser)
parser.add_argument(
"--port",
@@ -792,6 +822,8 @@ def create_recognizer(args) -> sherpa_onnx.OfflineRecognizer:
feature_dim=args.feat_dim,
decoding_method=args.decoding_method,
max_active_paths=args.max_active_paths,
hotwords_file=args.hotwords_file,
hotwords_score=args.hotwords_score,
)
elif args.paraformer:
assert len(args.nemo_ctc) == 0, args.nemo_ctc

72
python-api-examples/offline-decode-files.py
View File

@@ -82,7 +82,6 @@ from pathlib import Path
from typing import List, Tuple
import numpy as np
import sentencepiece as spm
import sherpa_onnx
@@ -98,43 +97,25 @@ def get_args():
)
parser.add_argument(
"--bpe-model",
"--hotwords-file",
type=str,
default="",
help="""
Path to bpe.model,
Used only when --decoding-method=modified_beam_search
The file containing hotwords, one words/phrases per line, and for each
phrase the bpe/cjkchar are separated by a space. For example:
▁HE LL O ▁WORLD
你 好 世 界
""",
)
parser.add_argument(
"--modeling-unit",
type=str,
default="char",
help="""
The type of modeling unit.
Valid values are bpe, bpe+char, char.
Note: the char here means characters in CJK languages.
""",
)
parser.add_argument(
"--contexts",
type=str,
default="",
help="""
The context list, it is a string containing some words/phrases separated
with /, for example, 'HELLO WORLD/I LOVE YOU/GO AWAY".
""",
)
parser.add_argument(
"--context-score",
"--hotwords-score",
type=float,
default=1.5,
help="""
The context score of each token for biasing word/phrase. Used only if
--contexts is given.
The hotword score of each token for biasing word/phrase. Used only if
--hotwords-file is given.
""",
)
@@ -273,25 +254,6 @@ def assert_file_exists(filename: str):
"https://k2-fsa.github.io/sherpa/onnx/pretrained_models/index.html to download it"
)
def encode_contexts(args, contexts: List[str]) -> List[List[int]]:
sp = None
if "bpe" in args.modeling_unit:
assert_file_exists(args.bpe_model)
sp = spm.SentencePieceProcessor()
sp.load(args.bpe_model)
tokens = {}
with open(args.tokens, "r", encoding="utf-8") as f:
for line in f:
toks = line.strip().split()
assert len(toks) == 2, len(toks)
assert toks[0] not in tokens, f"Duplicate token: {toks} "
tokens[toks[0]] = int(toks[1])
return sherpa_onnx.encode_contexts(
modeling_unit=args.modeling_unit, contexts=contexts, sp=sp, tokens_table=tokens
)
def read_wave(wave_filename: str) -> Tuple[np.ndarray, int]:
"""
Args:
@@ -322,7 +284,6 @@ def main():
assert_file_exists(args.tokens)
assert args.num_threads > 0, args.num_threads
contexts_list = []
if args.encoder:
assert len(args.paraformer) == 0, args.paraformer
assert len(args.nemo_ctc) == 0, args.nemo_ctc
@@ -330,11 +291,6 @@ def main():
assert len(args.whisper_decoder) == 0, args.whisper_decoder
assert len(args.tdnn_model) == 0, args.tdnn_model
contexts = [x.strip().upper() for x in args.contexts.split("/") if x.strip()]
if contexts:
print(f"Contexts list: {contexts}")
contexts_list = encode_contexts(args, contexts)
assert_file_exists(args.encoder)
assert_file_exists(args.decoder)
assert_file_exists(args.joiner)
@@ -348,7 +304,8 @@ def main():
sample_rate=args.sample_rate,
feature_dim=args.feature_dim,
decoding_method=args.decoding_method,
context_score=args.context_score,
hotwords_file=args.hotwords_file,
hotwords_score=args.hotwords_score,
debug=args.debug,
)
elif args.paraformer:
@@ -425,12 +382,7 @@ def main():
samples, sample_rate = read_wave(wave_filename)
duration = len(samples) / sample_rate
total_duration += duration
if contexts_list:
assert len(args.paraformer) == 0, args.paraformer
assert len(args.nemo_ctc) == 0, args.nemo_ctc
s = recognizer.create_stream(contexts_list=contexts_list)
else:
s = recognizer.create_stream()
s = recognizer.create_stream()
s.accept_waveform(sample_rate, samples)
streams.append(s)

75
python-api-examples/online-decode-files.py
View File

@@ -48,7 +48,6 @@ from pathlib import Path
from typing import List, Tuple
import numpy as np
import sentencepiece as spm
import sherpa_onnx
@@ -124,46 +123,25 @@ def get_args():
)
parser.add_argument(
"--bpe-model",
"--hotwords-file",
type=str,
default="",
help="""
Path to bpe.model, it will be used to tokenize contexts biasing phrases.
Used only when --decoding-method=modified_beam_search
The file containing hotwords, one words/phrases per line, and for each
phrase the bpe/cjkchar are separated by a space. For example:
▁HE LL O ▁WORLD
你 好 世 界
""",
)
parser.add_argument(
"--modeling-unit",
type=str,
default="char",
help="""
The type of modeling unit, it will be used to tokenize contexts biasing phrases.
Valid values are bpe, bpe+char, char.
Note: the char here means characters in CJK languages.
Used only when --decoding-method=modified_beam_search
""",
)
parser.add_argument(
"--contexts",
type=str,
default="",
help="""
The context list, it is a string containing some words/phrases separated
with /, for example, 'HELLO WORLD/I LOVE YOU/GO AWAY".
Used only when --decoding-method=modified_beam_search
""",
)
parser.add_argument(
"--context-score",
"--hotwords-score",
type=float,
default=1.5,
help="""
The context score of each token for biasing word/phrase. Used only if
--contexts is given.
Used only when --decoding-method=modified_beam_search
The hotword score of each token for biasing word/phrase. Used only if
--hotwords-file is given.
""",
)
@@ -214,27 +192,6 @@ def read_wave(wave_filename: str) -> Tuple[np.ndarray, int]:
return samples_float32, f.getframerate()
def encode_contexts(args, contexts: List[str]) -> List[List[int]]:
sp = None
if "bpe" in args.modeling_unit:
assert_file_exists(args.bpe_model)
sp = spm.SentencePieceProcessor()
sp.load(args.bpe_model)
tokens = {}
with open(args.tokens, "r", encoding="utf-8") as f:
for line in f:
toks = line.strip().split()
assert len(toks) == 2, len(toks)
assert toks[0] not in tokens, f"Duplicate token: {toks} "
tokens[toks[0]] = int(toks[1])
return sherpa_onnx.encode_contexts(
modeling_unit=args.modeling_unit,
contexts=contexts,
sp=sp,
tokens_table=tokens,
)
def main():
args = get_args()
assert_file_exists(args.tokens)
@@ -258,7 +215,8 @@ def main():
feature_dim=80,
decoding_method=args.decoding_method,
max_active_paths=args.max_active_paths,
context_score=args.context_score,
hotwords_file=args.hotwords_file,
hotwords_score=args.hotwords_score,
)
elif args.paraformer_encoder:
recognizer = sherpa_onnx.OnlineRecognizer.from_paraformer(
@@ -277,12 +235,6 @@ def main():
print("Started!")
start_time = time.time()
contexts_list = []
contexts = [x.strip().upper() for x in args.contexts.split("/") if x.strip()]
if contexts:
print(f"Contexts list: {contexts}")
contexts_list = encode_contexts(args, contexts)
streams = []
total_duration = 0
for wave_filename in args.sound_files:
@@ -291,10 +243,7 @@ def main():
duration = len(samples) / sample_rate
total_duration += duration
if contexts_list:
s = recognizer.create_stream(contexts_list=contexts_list)
else:
s = recognizer.create_stream()
s = recognizer.create_stream()
s.accept_waveform(sample_rate, samples)

26
python-api-examples/speech-recognition-from-microphone-with-endpoint-detection.py
View File

@@ -79,6 +79,30 @@ def get_args():
help="Valid values: cpu, cuda, coreml",
)
parser.add_argument(
"--hotwords-file",
type=str,
default="",
help="""
The file containing hotwords, one words/phrases per line, and for each
phrase the bpe/cjkchar are separated by a space. For example:
▁HE LL O ▁WORLD
你 好 世 界
""",
)
parser.add_argument(
"--hotwords-score",
type=float,
default=1.5,
help="""
The hotword score of each token for biasing word/phrase. Used only if
--hotwords-file is given.
""",
)
return parser.parse_args()
@@ -104,6 +128,8 @@ def create_recognizer(args):
rule3_min_utterance_length=300, # it essentially disables this rule
decoding_method=args.decoding_method,
provider=args.provider,
hotwords_file=agrs.hotwords_file,
hotwords_score=args.hotwords_score,
)
return recognizer

76
python-api-examples/speech-recognition-from-microphone.py
View File

@@ -11,7 +11,6 @@ import sys
from pathlib import Path
from typing import List
import sentencepiece as spm
try:
import sounddevice as sd
@@ -90,49 +89,29 @@ def get_args():
)
parser.add_argument(
"--bpe-model",
"--hotwords-file",
type=str,
default="",
help="""
Path to bpe.model, it will be used to tokenize contexts biasing phrases.
Used only when --decoding-method=modified_beam_search
The file containing hotwords, one words/phrases per line, and for each
phrase the bpe/cjkchar are separated by a space. For example:
▁HE LL O ▁WORLD
你 好 世 界
""",
)
parser.add_argument(
"--modeling-unit",
type=str,
default="char",
help="""
The type of modeling unit, it will be used to tokenize contexts biasing phrases.
Valid values are bpe, bpe+char, char.
Note: the char here means characters in CJK languages.
Used only when --decoding-method=modified_beam_search
""",
)
parser.add_argument(
"--contexts",
type=str,
default="",
help="""
The context list, it is a string containing some words/phrases separated
with /, for example, 'HELLO WORLD/I LOVE YOU/GO AWAY".
Used only when --decoding-method=modified_beam_search
""",
)
parser.add_argument(
"--context-score",
"--hotwords-score",
type=float,
default=1.5,
help="""
The context score of each token for biasing word/phrase. Used only if
--contexts is given.
Used only when --decoding-method=modified_beam_search
The hotword score of each token for biasing word/phrase. Used only if
--hotwords-file is given.
""",
)
return parser.parse_args()
@@ -155,32 +134,12 @@ def create_recognizer(args):
decoding_method=args.decoding_method,
max_active_paths=args.max_active_paths,
provider=args.provider,
context_score=args.context_score,
hotwords_file=args.hotwords_file,
hotwords_score=args.hotwords_score,
)
return recognizer
def encode_contexts(args, contexts: List[str]) -> List[List[int]]:
sp = None
if "bpe" in args.modeling_unit:
assert_file_exists(args.bpe_model)
sp = spm.SentencePieceProcessor()
sp.load(args.bpe_model)
tokens = {}
with open(args.tokens, "r", encoding="utf-8") as f:
for line in f:
toks = line.strip().split()
assert len(toks) == 2, len(toks)
assert toks[0] not in tokens, f"Duplicate token: {toks} "
tokens[toks[0]] = int(toks[1])
return sherpa_onnx.encode_contexts(
modeling_unit=args.modeling_unit,
contexts=contexts,
sp=sp,
tokens_table=tokens,
)
def main():
args = get_args()
@@ -193,12 +152,6 @@ def main():
default_input_device_idx = sd.default.device[0]
print(f'Use default device: {devices[default_input_device_idx]["name"]}')
contexts_list = []
contexts = [x.strip().upper() for x in args.contexts.split("/") if x.strip()]
if contexts:
print(f"Contexts list: {contexts}")
contexts_list = encode_contexts(args, contexts)
recognizer = create_recognizer(args)
print("Started! Please speak")
@@ -207,10 +160,7 @@ def main():
sample_rate = 48000
samples_per_read = int(0.1 * sample_rate) # 0.1 second = 100 ms
last_result = ""
if contexts_list:
stream = recognizer.create_stream(contexts_list=contexts_list)
else:
stream = recognizer.create_stream()
stream = recognizer.create_stream()
with sd.InputStream(channels=1, dtype="float32", samplerate=sample_rate) as s:
while True:
samples, _ = s.read(samples_per_read) # a blocking read

26
python-api-examples/speech-recognition-from-url.py
View File

@@ -87,6 +87,30 @@ def get_args():
""",
)
parser.add_argument(
"--hotwords-file",
type=str,
default="",
help="""
The file containing hotwords, one words/phrases per line, and for each
phrase the bpe/cjkchar are separated by a space. For example:
▁HE LL O ▁WORLD
你 好 世 界
""",
)
parser.add_argument(
"--hotwords-score",
type=float,
default=1.5,
help="""
The hotword score of each token for biasing word/phrase. Used only if
--hotwords-file is given.
""",
)
return parser.parse_args()
@@ -107,6 +131,8 @@ def create_recognizer(args):
rule1_min_trailing_silence=2.4,
rule2_min_trailing_silence=1.2,
rule3_min_utterance_length=300, # it essentially disables this rule
hotwords_file=args.hotwords_file,
hotwords_score=args.hotwords_score,
)
return recognizer

29
python-api-examples/streaming_server.py
View File

@@ -187,6 +187,32 @@ def add_decoding_args(parser: argparse.ArgumentParser):
add_modified_beam_search_args(parser)
def add_hotwords_args(parser: argparse.ArgumentParser):
parser.add_argument(
"--hotwords-file",
type=str,
default="",
help="""
The file containing hotwords, one words/phrases per line, and for each
phrase the bpe/cjkchar are separated by a space. For example:
▁HE LL O ▁WORLD
你 好 世 界
""",
)
parser.add_argument(
"--hotwords-score",
type=float,
default=1.5,
help="""
The hotword score of each token for biasing word/phrase. Used only if
--hotwords-file is given.
""",
)
def add_modified_beam_search_args(parser: argparse.ArgumentParser):
parser.add_argument(
"--num-active-paths",
@@ -239,6 +265,7 @@ def get_args():
add_model_args(parser)
add_decoding_args(parser)
add_endpointing_args(parser)
add_hotwords_args(parser)
parser.add_argument(
"--port",
@@ -343,6 +370,8 @@ def create_recognizer(args) -> sherpa_onnx.OnlineRecognizer:
feature_dim=args.feat_dim,
decoding_method=args.decoding_method,
max_active_paths=args.num_active_paths,
hotwords_score=args.hotwords_score,
hotwords_file=args.hotwords_file,
enable_endpoint_detection=args.use_endpoint != 0,
rule1_min_trailing_silence=args.rule1_min_trailing_silence,
rule2_min_trailing_silence=args.rule2_min_trailing_silence,