2e45d327a5
* Adding temperature scaling on Joiner logits:
- T hard-coded to 2.0
- so far best result NCE 0.122 (still not so high)
- the BPE scores were rescaled with 0.2 (but then also incorrect words
get high confidence, visually reasonable histograms are for 0.5 scale)
- BPE->WORD score merging done by min(.) function
(tried also prob-product, and also arithmetic, geometric, harmonic mean)
- without temperature scaling (i.e. scale 1.0), the best NCE was 0.032 (here product merging was best)
Results seem consistent with: https://arxiv.org/abs/2110.15222
Everything tuned on a very-small set of 100 sentences with 813 words and 10.2% WER, a Czech model.
I also experimented with blank posteriors mixed into the BPE confidences,
but no NCE improvement found, so not pushing that.
Temperature scling added also to the Greedy search confidences.
* making `temperature_scale` configurable from outside
569 lines
20 KiB
Python
569 lines
20 KiB
Python
# Copyright (c) 2023 Xiaomi Corporation
|
|
from pathlib import Path
|
|
from typing import List, Optional
|
|
|
|
from _sherpa_onnx import (
|
|
EndpointConfig,
|
|
FeatureExtractorConfig,
|
|
OnlineLMConfig,
|
|
OnlineModelConfig,
|
|
OnlineParaformerModelConfig,
|
|
)
|
|
from _sherpa_onnx import OnlineRecognizer as _Recognizer
|
|
from _sherpa_onnx import (
|
|
OnlineRecognizerConfig,
|
|
OnlineStream,
|
|
OnlineTransducerModelConfig,
|
|
OnlineWenetCtcModelConfig,
|
|
OnlineZipformer2CtcModelConfig,
|
|
OnlineCtcFstDecoderConfig,
|
|
)
|
|
|
|
|
|
def _assert_file_exists(f: str):
|
|
assert Path(f).is_file(), f"{f} does not exist"
|
|
|
|
|
|
class OnlineRecognizer(object):
|
|
"""A class for streaming speech recognition.
|
|
|
|
Please refer to the following files for usages
|
|
- https://github.com/k2-fsa/sherpa-onnx/blob/master/sherpa-onnx/python/tests/test_online_recognizer.py
|
|
- https://github.com/k2-fsa/sherpa-onnx/blob/master/python-api-examples/online-decode-files.py
|
|
"""
|
|
|
|
@classmethod
|
|
def from_transducer(
|
|
cls,
|
|
tokens: str,
|
|
encoder: str,
|
|
decoder: str,
|
|
joiner: str,
|
|
num_threads: int = 2,
|
|
sample_rate: float = 16000,
|
|
feature_dim: int = 80,
|
|
low_freq: float = 20.0,
|
|
high_freq: float = -400.0,
|
|
dither: float = 0.0,
|
|
enable_endpoint_detection: bool = False,
|
|
rule1_min_trailing_silence: float = 2.4,
|
|
rule2_min_trailing_silence: float = 1.2,
|
|
rule3_min_utterance_length: float = 20.0,
|
|
decoding_method: str = "greedy_search",
|
|
max_active_paths: int = 4,
|
|
hotwords_score: float = 1.5,
|
|
blank_penalty: float = 0.0,
|
|
hotwords_file: str = "",
|
|
provider: str = "cpu",
|
|
model_type: str = "",
|
|
lm: str = "",
|
|
lm_scale: float = 0.1,
|
|
temperature_scale: float = 2.0,
|
|
):
|
|
"""
|
|
Please refer to
|
|
`<https://k2-fsa.github.io/sherpa/onnx/pretrained_models/index.html>`_
|
|
to download pre-trained models for different languages, e.g., Chinese,
|
|
English, etc.
|
|
|
|
Args:
|
|
tokens:
|
|
Path to ``tokens.txt``. Each line in ``tokens.txt`` contains two
|
|
columns::
|
|
|
|
symbol integer_id
|
|
|
|
encoder:
|
|
Path to ``encoder.onnx``.
|
|
decoder:
|
|
Path to ``decoder.onnx``.
|
|
joiner:
|
|
Path to ``joiner.onnx``.
|
|
num_threads:
|
|
Number of threads for neural network computation.
|
|
sample_rate:
|
|
Sample rate of the training data used to train the model.
|
|
feature_dim:
|
|
Dimension of the feature used to train the model.
|
|
low_freq:
|
|
Low cutoff frequency for mel bins in feature extraction.
|
|
high_freq:
|
|
High cutoff frequency for mel bins in feature extraction
|
|
(if <= 0, offset from Nyquist)
|
|
dither:
|
|
Dithering constant (0.0 means no dither).
|
|
By default the audio samples are in range [-1,+1],
|
|
so dithering constant 0.00003 is a good value,
|
|
equivalent to the default 1.0 from kaldi
|
|
enable_endpoint_detection:
|
|
True to enable endpoint detection. False to disable endpoint
|
|
detection.
|
|
rule1_min_trailing_silence:
|
|
Used only when enable_endpoint_detection is True. If the duration
|
|
of trailing silence in seconds is larger than this value, we assume
|
|
an endpoint is detected.
|
|
rule2_min_trailing_silence:
|
|
Used only when enable_endpoint_detection is True. If we have decoded
|
|
something that is nonsilence and if the duration of trailing silence
|
|
in seconds is larger than this value, we assume an endpoint is
|
|
detected.
|
|
rule3_min_utterance_length:
|
|
Used only when enable_endpoint_detection is True. If the utterance
|
|
length in seconds is larger than this value, we assume an endpoint
|
|
is detected.
|
|
decoding_method:
|
|
Valid values are greedy_search, modified_beam_search.
|
|
max_active_paths:
|
|
Use only when decoding_method is modified_beam_search. It specifies
|
|
the maximum number of active paths during beam search.
|
|
blank_penalty:
|
|
The penalty applied on blank symbol during decoding.
|
|
hotwords_file:
|
|
The file containing hotwords, one words/phrases per line, and for each
|
|
phrase the bpe/cjkchar are separated by a space.
|
|
hotwords_score:
|
|
The hotword score of each token for biasing word/phrase. Used only if
|
|
hotwords_file is given with modified_beam_search as decoding method.
|
|
temperature_scale:
|
|
Temperature scaling for output symbol confidence estiamation.
|
|
It affects only confidence values, the decoding uses the original
|
|
logits without temperature.
|
|
provider:
|
|
onnxruntime execution providers. Valid values are: cpu, cuda, coreml.
|
|
model_type:
|
|
Online transducer model type. Valid values are: conformer, lstm,
|
|
zipformer, zipformer2. All other values lead to loading the model twice.
|
|
"""
|
|
self = cls.__new__(cls)
|
|
_assert_file_exists(tokens)
|
|
_assert_file_exists(encoder)
|
|
_assert_file_exists(decoder)
|
|
_assert_file_exists(joiner)
|
|
|
|
assert num_threads > 0, num_threads
|
|
|
|
transducer_config = OnlineTransducerModelConfig(
|
|
encoder=encoder,
|
|
decoder=decoder,
|
|
joiner=joiner,
|
|
)
|
|
|
|
model_config = OnlineModelConfig(
|
|
transducer=transducer_config,
|
|
tokens=tokens,
|
|
num_threads=num_threads,
|
|
provider=provider,
|
|
model_type=model_type,
|
|
)
|
|
|
|
feat_config = FeatureExtractorConfig(
|
|
sampling_rate=sample_rate,
|
|
feature_dim=feature_dim,
|
|
low_freq=low_freq,
|
|
high_freq=high_freq,
|
|
dither=dither,
|
|
)
|
|
|
|
endpoint_config = EndpointConfig(
|
|
rule1_min_trailing_silence=rule1_min_trailing_silence,
|
|
rule2_min_trailing_silence=rule2_min_trailing_silence,
|
|
rule3_min_utterance_length=rule3_min_utterance_length,
|
|
)
|
|
|
|
if len(hotwords_file) > 0 and decoding_method != "modified_beam_search":
|
|
raise ValueError(
|
|
"Please use --decoding-method=modified_beam_search when using "
|
|
f"--hotwords-file. Currently given: {decoding_method}"
|
|
)
|
|
|
|
if lm and decoding_method != "modified_beam_search":
|
|
raise ValueError(
|
|
"Please use --decoding-method=modified_beam_search when using "
|
|
f"--lm. Currently given: {decoding_method}"
|
|
)
|
|
|
|
lm_config = OnlineLMConfig(
|
|
model=lm,
|
|
scale=lm_scale,
|
|
)
|
|
|
|
recognizer_config = OnlineRecognizerConfig(
|
|
feat_config=feat_config,
|
|
model_config=model_config,
|
|
lm_config=lm_config,
|
|
endpoint_config=endpoint_config,
|
|
enable_endpoint=enable_endpoint_detection,
|
|
decoding_method=decoding_method,
|
|
max_active_paths=max_active_paths,
|
|
hotwords_score=hotwords_score,
|
|
hotwords_file=hotwords_file,
|
|
blank_penalty=blank_penalty,
|
|
temperature_scale=temperature_scale,
|
|
)
|
|
|
|
self.recognizer = _Recognizer(recognizer_config)
|
|
self.config = recognizer_config
|
|
return self
|
|
|
|
@classmethod
|
|
def from_paraformer(
|
|
cls,
|
|
tokens: str,
|
|
encoder: str,
|
|
decoder: str,
|
|
num_threads: int = 2,
|
|
sample_rate: float = 16000,
|
|
feature_dim: int = 80,
|
|
enable_endpoint_detection: bool = False,
|
|
rule1_min_trailing_silence: float = 2.4,
|
|
rule2_min_trailing_silence: float = 1.2,
|
|
rule3_min_utterance_length: float = 20.0,
|
|
decoding_method: str = "greedy_search",
|
|
provider: str = "cpu",
|
|
):
|
|
"""
|
|
Please refer to
|
|
`<https://k2-fsa.github.io/sherpa/onnx/pretrained_models/index.html>`_
|
|
to download pre-trained models for different languages, e.g., Chinese,
|
|
English, etc.
|
|
|
|
Args:
|
|
tokens:
|
|
Path to ``tokens.txt``. Each line in ``tokens.txt`` contains two
|
|
columns::
|
|
|
|
symbol integer_id
|
|
|
|
encoder:
|
|
Path to ``encoder.onnx``.
|
|
decoder:
|
|
Path to ``decoder.onnx``.
|
|
num_threads:
|
|
Number of threads for neural network computation.
|
|
sample_rate:
|
|
Sample rate of the training data used to train the model.
|
|
feature_dim:
|
|
Dimension of the feature used to train the model.
|
|
enable_endpoint_detection:
|
|
True to enable endpoint detection. False to disable endpoint
|
|
detection.
|
|
rule1_min_trailing_silence:
|
|
Used only when enable_endpoint_detection is True. If the duration
|
|
of trailing silence in seconds is larger than this value, we assume
|
|
an endpoint is detected.
|
|
rule2_min_trailing_silence:
|
|
Used only when enable_endpoint_detection is True. If we have decoded
|
|
something that is nonsilence and if the duration of trailing silence
|
|
in seconds is larger than this value, we assume an endpoint is
|
|
detected.
|
|
rule3_min_utterance_length:
|
|
Used only when enable_endpoint_detection is True. If the utterance
|
|
length in seconds is larger than this value, we assume an endpoint
|
|
is detected.
|
|
decoding_method:
|
|
The only valid value is greedy_search.
|
|
provider:
|
|
onnxruntime execution providers. Valid values are: cpu, cuda, coreml.
|
|
"""
|
|
self = cls.__new__(cls)
|
|
_assert_file_exists(tokens)
|
|
_assert_file_exists(encoder)
|
|
_assert_file_exists(decoder)
|
|
|
|
assert num_threads > 0, num_threads
|
|
|
|
paraformer_config = OnlineParaformerModelConfig(
|
|
encoder=encoder,
|
|
decoder=decoder,
|
|
)
|
|
|
|
model_config = OnlineModelConfig(
|
|
paraformer=paraformer_config,
|
|
tokens=tokens,
|
|
num_threads=num_threads,
|
|
provider=provider,
|
|
model_type="paraformer",
|
|
)
|
|
|
|
feat_config = FeatureExtractorConfig(
|
|
sampling_rate=sample_rate,
|
|
feature_dim=feature_dim,
|
|
)
|
|
|
|
endpoint_config = EndpointConfig(
|
|
rule1_min_trailing_silence=rule1_min_trailing_silence,
|
|
rule2_min_trailing_silence=rule2_min_trailing_silence,
|
|
rule3_min_utterance_length=rule3_min_utterance_length,
|
|
)
|
|
|
|
recognizer_config = OnlineRecognizerConfig(
|
|
feat_config=feat_config,
|
|
model_config=model_config,
|
|
endpoint_config=endpoint_config,
|
|
enable_endpoint=enable_endpoint_detection,
|
|
decoding_method=decoding_method,
|
|
)
|
|
|
|
self.recognizer = _Recognizer(recognizer_config)
|
|
self.config = recognizer_config
|
|
return self
|
|
|
|
@classmethod
|
|
def from_zipformer2_ctc(
|
|
cls,
|
|
tokens: str,
|
|
model: str,
|
|
num_threads: int = 2,
|
|
sample_rate: float = 16000,
|
|
feature_dim: int = 80,
|
|
enable_endpoint_detection: bool = False,
|
|
rule1_min_trailing_silence: float = 2.4,
|
|
rule2_min_trailing_silence: float = 1.2,
|
|
rule3_min_utterance_length: float = 20.0,
|
|
decoding_method: str = "greedy_search",
|
|
ctc_graph: str = "",
|
|
ctc_max_active: int = 3000,
|
|
provider: str = "cpu",
|
|
):
|
|
"""
|
|
Please refer to
|
|
`<https://k2-fsa.github.io/sherpa/onnx/pretrained_models/online-ctc/index.html>`_
|
|
to download pre-trained models for different languages, e.g., Chinese,
|
|
English, etc.
|
|
|
|
Args:
|
|
tokens:
|
|
Path to ``tokens.txt``. Each line in ``tokens.txt`` contains two
|
|
columns::
|
|
|
|
symbol integer_id
|
|
|
|
model:
|
|
Path to ``model.onnx``.
|
|
num_threads:
|
|
Number of threads for neural network computation.
|
|
sample_rate:
|
|
Sample rate of the training data used to train the model.
|
|
feature_dim:
|
|
Dimension of the feature used to train the model.
|
|
enable_endpoint_detection:
|
|
True to enable endpoint detection. False to disable endpoint
|
|
detection.
|
|
rule1_min_trailing_silence:
|
|
Used only when enable_endpoint_detection is True. If the duration
|
|
of trailing silence in seconds is larger than this value, we assume
|
|
an endpoint is detected.
|
|
rule2_min_trailing_silence:
|
|
Used only when enable_endpoint_detection is True. If we have decoded
|
|
something that is nonsilence and if the duration of trailing silence
|
|
in seconds is larger than this value, we assume an endpoint is
|
|
detected.
|
|
rule3_min_utterance_length:
|
|
Used only when enable_endpoint_detection is True. If the utterance
|
|
length in seconds is larger than this value, we assume an endpoint
|
|
is detected.
|
|
decoding_method:
|
|
The only valid value is greedy_search.
|
|
ctc_graph:
|
|
If not empty, decoding_method is ignored. It contains the path to
|
|
H.fst, HL.fst, or HLG.fst
|
|
ctc_max_active:
|
|
Used only when ctc_graph is not empty. It specifies the maximum
|
|
active paths at a time.
|
|
provider:
|
|
onnxruntime execution providers. Valid values are: cpu, cuda, coreml.
|
|
"""
|
|
self = cls.__new__(cls)
|
|
_assert_file_exists(tokens)
|
|
_assert_file_exists(model)
|
|
|
|
assert num_threads > 0, num_threads
|
|
|
|
zipformer2_ctc_config = OnlineZipformer2CtcModelConfig(model=model)
|
|
|
|
model_config = OnlineModelConfig(
|
|
zipformer2_ctc=zipformer2_ctc_config,
|
|
tokens=tokens,
|
|
num_threads=num_threads,
|
|
provider=provider,
|
|
)
|
|
|
|
feat_config = FeatureExtractorConfig(
|
|
sampling_rate=sample_rate,
|
|
feature_dim=feature_dim,
|
|
)
|
|
|
|
endpoint_config = EndpointConfig(
|
|
rule1_min_trailing_silence=rule1_min_trailing_silence,
|
|
rule2_min_trailing_silence=rule2_min_trailing_silence,
|
|
rule3_min_utterance_length=rule3_min_utterance_length,
|
|
)
|
|
|
|
ctc_fst_decoder_config = OnlineCtcFstDecoderConfig(
|
|
graph=ctc_graph,
|
|
max_active=ctc_max_active,
|
|
)
|
|
|
|
recognizer_config = OnlineRecognizerConfig(
|
|
feat_config=feat_config,
|
|
model_config=model_config,
|
|
endpoint_config=endpoint_config,
|
|
ctc_fst_decoder_config=ctc_fst_decoder_config,
|
|
enable_endpoint=enable_endpoint_detection,
|
|
decoding_method=decoding_method,
|
|
)
|
|
|
|
self.recognizer = _Recognizer(recognizer_config)
|
|
self.config = recognizer_config
|
|
return self
|
|
|
|
@classmethod
|
|
def from_wenet_ctc(
|
|
cls,
|
|
tokens: str,
|
|
model: str,
|
|
chunk_size: int = 16,
|
|
num_left_chunks: int = 4,
|
|
num_threads: int = 2,
|
|
sample_rate: float = 16000,
|
|
feature_dim: int = 80,
|
|
enable_endpoint_detection: bool = False,
|
|
rule1_min_trailing_silence: float = 2.4,
|
|
rule2_min_trailing_silence: float = 1.2,
|
|
rule3_min_utterance_length: float = 20.0,
|
|
decoding_method: str = "greedy_search",
|
|
provider: str = "cpu",
|
|
):
|
|
"""
|
|
Please refer to
|
|
`<https://k2-fsa.github.io/sherpa/onnx/pretrained_models/wenet/index.html>`_
|
|
to download pre-trained models for different languages, e.g., Chinese,
|
|
English, etc.
|
|
|
|
Args:
|
|
tokens:
|
|
Path to ``tokens.txt``. Each line in ``tokens.txt`` contains two
|
|
columns::
|
|
|
|
symbol integer_id
|
|
|
|
model:
|
|
Path to ``model.onnx``.
|
|
chunk_size:
|
|
The --chunk-size parameter from WeNet.
|
|
num_left_chunks:
|
|
The --num-left-chunks parameter from WeNet.
|
|
num_threads:
|
|
Number of threads for neural network computation.
|
|
sample_rate:
|
|
Sample rate of the training data used to train the model.
|
|
feature_dim:
|
|
Dimension of the feature used to train the model.
|
|
enable_endpoint_detection:
|
|
True to enable endpoint detection. False to disable endpoint
|
|
detection.
|
|
rule1_min_trailing_silence:
|
|
Used only when enable_endpoint_detection is True. If the duration
|
|
of trailing silence in seconds is larger than this value, we assume
|
|
an endpoint is detected.
|
|
rule2_min_trailing_silence:
|
|
Used only when enable_endpoint_detection is True. If we have decoded
|
|
something that is nonsilence and if the duration of trailing silence
|
|
in seconds is larger than this value, we assume an endpoint is
|
|
detected.
|
|
rule3_min_utterance_length:
|
|
Used only when enable_endpoint_detection is True. If the utterance
|
|
length in seconds is larger than this value, we assume an endpoint
|
|
is detected.
|
|
decoding_method:
|
|
The only valid value is greedy_search.
|
|
provider:
|
|
onnxruntime execution providers. Valid values are: cpu, cuda, coreml.
|
|
"""
|
|
self = cls.__new__(cls)
|
|
_assert_file_exists(tokens)
|
|
_assert_file_exists(model)
|
|
|
|
assert num_threads > 0, num_threads
|
|
|
|
wenet_ctc_config = OnlineWenetCtcModelConfig(
|
|
model=model,
|
|
chunk_size=chunk_size,
|
|
num_left_chunks=num_left_chunks,
|
|
)
|
|
|
|
model_config = OnlineModelConfig(
|
|
wenet_ctc=wenet_ctc_config,
|
|
tokens=tokens,
|
|
num_threads=num_threads,
|
|
provider=provider,
|
|
)
|
|
|
|
feat_config = FeatureExtractorConfig(
|
|
sampling_rate=sample_rate,
|
|
feature_dim=feature_dim,
|
|
)
|
|
|
|
endpoint_config = EndpointConfig(
|
|
rule1_min_trailing_silence=rule1_min_trailing_silence,
|
|
rule2_min_trailing_silence=rule2_min_trailing_silence,
|
|
rule3_min_utterance_length=rule3_min_utterance_length,
|
|
)
|
|
|
|
recognizer_config = OnlineRecognizerConfig(
|
|
feat_config=feat_config,
|
|
model_config=model_config,
|
|
endpoint_config=endpoint_config,
|
|
enable_endpoint=enable_endpoint_detection,
|
|
decoding_method=decoding_method,
|
|
)
|
|
|
|
self.recognizer = _Recognizer(recognizer_config)
|
|
self.config = recognizer_config
|
|
return self
|
|
|
|
def create_stream(self, hotwords: Optional[str] = None):
|
|
if hotwords is None:
|
|
return self.recognizer.create_stream()
|
|
else:
|
|
return self.recognizer.create_stream(hotwords)
|
|
|
|
def decode_stream(self, s: OnlineStream):
|
|
self.recognizer.decode_stream(s)
|
|
|
|
def decode_streams(self, ss: List[OnlineStream]):
|
|
self.recognizer.decode_streams(ss)
|
|
|
|
def is_ready(self, s: OnlineStream) -> bool:
|
|
return self.recognizer.is_ready(s)
|
|
|
|
def get_result(self, s: OnlineStream) -> str:
|
|
return self.recognizer.get_result(s).text.strip()
|
|
|
|
def get_result_as_json_string(self, s: OnlineStream) -> str:
|
|
return self.recognizer.get_result(s).as_json_string()
|
|
|
|
def tokens(self, s: OnlineStream) -> List[str]:
|
|
return self.recognizer.get_result(s).tokens
|
|
|
|
def timestamps(self, s: OnlineStream) -> List[float]:
|
|
return self.recognizer.get_result(s).timestamps
|
|
|
|
def start_time(self, s: OnlineStream) -> float:
|
|
return self.recognizer.get_result(s).start_time
|
|
|
|
def ys_probs(self, s: OnlineStream) -> List[float]:
|
|
return self.recognizer.get_result(s).ys_probs
|
|
|
|
def lm_probs(self, s: OnlineStream) -> List[float]:
|
|
return self.recognizer.get_result(s).lm_probs
|
|
|
|
def context_scores(self, s: OnlineStream) -> List[float]:
|
|
return self.recognizer.get_result(s).context_scores
|
|
|
|
def is_endpoint(self, s: OnlineStream) -> bool:
|
|
return self.recognizer.is_endpoint(s)
|
|
|
|
def reset(self, s: OnlineStream) -> bool:
|
|
return self.recognizer.reset(s)
|