-cJjddlmZddlmZmZddlmZmZmZm Z m Z m Z m Z m Z mZddlmZddlmZmZmZmZmZmZmZmZmZdefdZGdd eZy ) )Path)ListOptional) CudaConfigEndpointConfigFeatureExtractorConfigHomophoneReplacerConfigOnlineCtcFstDecoderConfigOnlineLMConfigOnlineModelConfigOnlineNeMoCtcModelConfigOnlineParaformerModelConfig)OnlineRecognizer) OnlineRecognizerConfigOnlineRecognizerResult OnlineStreamOnlineToneCtcModelConfigOnlineTransducerModelConfigOnlineWenetCtcModelConfigOnlineZipformer2CtcModelConfigProviderConfigTensorrtConfigfcJt|js J|dy)Nz does not exist)ris_file)rs g/Users/ahmed/devFolder/claude-voice/.venv/lib/python3.12/site-packages/sherpa_onnx/online_recognizer.py_assert_file_existsrs" 7?? 3?33 cfeZdZdZe dSdededededededed ed ed ed ed ededededededededededededededededededed ed!ed"ed#ed$ed%ed&ed'ed(ed)ed*ed+ed,ed-ed.ed/ed0ed1ed2ed3efdd4Z e dTdededededededededededed"eded ed!ed#ed/ed0ed1ef&d5Z e dUded6ededededededededed7ed8ed"eded ed!ed#ed/ed0ed1ef(d9Z e dVded6ededededededededed"eded ed!ed#ed/ed0ed1ef$d:Z e dTded6ededededededededed"eded ed!ed#ed/ed0ed1ef$d;Z e dWded6edZdXd@eefdAZdBefdCZdDeefdEZdBedFefdGZdBedFefdHZdBedFefdIZdBedFefdJZdBedFeefdKZdBedFeefdLZdBedFefdMZdBedFeefdNZdBedFeefdOZdBedFeefdPZdBedFefdQZ dBedFefdRZ!y?)YraA class for streaming (online) speech recognition. It supports multiple model families via factory methods: - :meth:`from_transducer` -- Zipformer, Nemotron, etc. - :meth:`from_paraformer` -- Streaming Paraformer - :meth:`from_zipformer2_ctc` -- Zipformer2 CTC - :meth:`from_wenet_ctc` -- WeNet CTC Example using streaming Zipformer transducer:: import numpy as np import sherpa_onnx recognizer = sherpa_onnx.OnlineRecognizer.from_transducer( tokens="./model/tokens.txt", encoder="./model/encoder-epoch-99-avg-1.int8.onnx", decoder="./model/decoder-epoch-99-avg-1.onnx", joiner="./model/joiner-epoch-99-avg-1.int8.onnx", num_threads=2, decoding_method="greedy_search", enable_endpoint_detection=True, ) # Read audio (float32, normalized to [-1, 1]) samples = read_audio("test.wav") # your own function sample_rate = 16000 stream = recognizer.create_stream() stream.accept_waveform(sample_rate, samples) # Add tail padding and signal end of input tail = np.zeros(int(0.5 * sample_rate), dtype=np.float32) stream.accept_waveform(sample_rate, tail) stream.input_finished() # Streaming decode loop while recognizer.is_ready(stream): recognizer.decode_stream(stream) result = recognizer.get_result_all(stream) print(result.text) Example with endpoint detection:: stream = recognizer.create_stream() stream.accept_waveform(sample_rate, samples) while recognizer.is_ready(stream): recognizer.decode_stream(stream) if recognizer.is_endpoint(stream): result = recognizer.get_result_all(stream) print("Endpoint detected:", result.text) recognizer.reset(stream) # reset for next utterance Please refer to the following files for more usages: - ``_ - ``_ tokensencoderdecoderjoiner num_threads sample_rate feature_dimlow_freq high_freqdithernormalize_samples snip_edgesenable_endpoint_detectionrule1_min_trailing_silencerule2_min_trailing_silencerule3_min_utterance_lengthdecoding_methodmax_active_pathshotwords_score blank_penalty hotwords_file model_type modeling_unit bpe_vocablmlm_scalelm_shallow_fusiontemperature_scale reset_encoderdebug rule_fsts rule_farsproviderdevicecudnn_conv_algo_searchtrt_max_workspace_sizetrt_max_partition_iterationstrt_min_subgraph_sizetrt_fp16_enabletrt_detailed_build_logtrt_engine_cache_enabletrt_timing_cache_enabletrt_engine_cache_pathtrt_timing_cache_pathtrt_dump_subgraphs hr_dict_dir hr_rule_fsts hr_lexiconlodr_fst lodr_scalec3h|j|}3t|t|t|t|t|||}4t|#}5t |$|%|&|'|(|)|*|+|,|- }6t |6|5|!|"}7t |4|||7||||}8t|| | ||| | }9t|||}:t|dkDr|d k7rtd ||r|d k7rtd |t||||1|2 };t|9|8|;|:| |||||||| |t|.|0|/ }`_ 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 normalize_samples: True for +/- 1.0 range of audio samples (default, zipformer feats), False for +/- 32k samples (ebranchformer features). snip_edges: handling of end of audio signal in kaldi feature extraction. If true, end effects will be handled by outputting only frames that completely fit in the file, and the number of frames depends on the frame-length. If false, the number of frames depends only on the frame-shift, and we reflect the data at the ends. 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 estimation. It affects only confidence values, the decoding uses the original logits without temperature. reset_encoder: True to reset `encoder_state` on an endpoint after empty segment. Done in `Reset()` method, after an endpoint was detected, currently only in `OnlineRecognizerTransducerImpl`. model_type: Online transducer model type. Valid values are: conformer, lstm, zipformer, zipformer2. All other values lead to loading the model twice. modeling_unit: The modeling unit of the model, commonly used units are bpe, cjkchar, cjkchar+bpe, etc. Currently, it is needed only when hotwords are provided, we need it to encode the hotwords into token sequence. bpe_vocab: The vocabulary generated by google's sentencepiece program. It is a file has two columns, one is the token, the other is the log probability, you can get it from the directory where your bpe model is generated. Only used when hotwords provided and the modeling unit is bpe or cjkchar+bpe. rule_fsts: If not empty, it specifies fsts for inverse text normalization. If there are multiple fsts, they are separated by a comma. rule_fars: If not empty, it specifies fst archives for inverse text normalization. If there are multiple archives, they are separated by a comma. provider: onnxruntime execution providers. Valid values are: cpu, cuda, coreml. device: onnxruntime cuda device index. cudnn_conv_algo_search: onxrt CuDNN convolution search algorithm selection. CUDA EP trt_max_workspace_size: Set TensorRT EP GPU memory usage limit. TensorRT EP trt_max_partition_iterations: Limit partitioning iterations for model conversion. TensorRT EP trt_min_subgraph_size: Set minimum size for subgraphs in partitioning. TensorRT EP trt_fp16_enable: bool = True, Enable FP16 precision for faster performance. TensorRT EP trt_detailed_build_log: bool = False, Enable detailed logging of build steps. TensorRT EP trt_engine_cache_enable: bool = True, Enable caching of TensorRT engines. TensorRT EP trt_timing_cache_enable: bool = True, "Enable use of timing cache to speed up builds." TensorRT EP trt_engine_cache_path: str ="", "Set path to store cached TensorRT engines." TensorRT EP trt_timing_cache_path: str ="", "Set path for storing timing cache." TensorRT EP trt_dump_subgraphs: bool = False, "Dump optimized subgraphs for debugging." TensorRT EP lodr_fst: Path to the LODR FST file in binary format. If empty, LODR is disabled. lodr_scale: Scale factor for LODR rescoring. Only used when lodr_fst is provided. )r!r"r#)rB) rCrDrErFrGrHrIrJrKrL) trt_config cuda_configr@rA) transducerr r$provider_configr5r6r7r=) sampling_rater*r+r&r'r(r)r-r.r/rmodified_beam_searchz_Please use --decoding-method=modified_beam_search when using --hotwords-file. Currently given: zTPlease use --decoding-method=modified_beam_search when using --lm. Currently given: )modelscaleshallow_fusionrPrQdict_dirlexiconr>) feat_config model_config lm_configendpoint_configenable_endpointr0r1r2r4r3r;r>r?r<hr)__new__rrrrrr rrlen ValueErrorr rr _Recognizer recognizerconfig)=clsr r!r"r#r$r%r&r'r(r)r*r+r,r-r.r/r0r1r2r3r4r5r6r7r8r9r:r;r<r=r>r?r@rArBrCrDrErFrGrHrIrJrKrLrMrNrOrPrQselftransducer_configrTrSrVrar`rcrbrecognizer_configs= rfrom_transducerz OnlineRecognizer.from_transducerasr{{3F#G$G$F#7  !#9 $#9)E"7+#9$;$;"7"71  )!#  )(#+!'  -%/!# )'A'A'A  }  !o9O&O55D4EG  /%;;**9):<  #,!  3#%+5+-)''/'&$"& ,&&78'  rcb|j|}t|t|t|t||}t| |}t ||||d| }t ||}t || | }t||||| ||t|||}t||_ ||_ |S) a Please refer to ``_ 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. rule_fsts: If not empty, it specifies fsts for inverse text normalization. If there are multiple fsts, they are separated by a comma. rule_fars: If not empty, it specifies fst archives for inverse text normalization. If there are multiple archives, they are separated by a comma. device: onnxruntime cuda device index. )r!r"r@rA paraformer)rsr r$rVr5r=rWr&rXr]r`rarcrdr0r>r?re) rfrrrr rrrr rirjrk)rlr r!r"r$r%r&r,r-r.r/r0r@r=r>r?rArMrNrOrmparaformer_configrVrar`rcros rfrom_paraformerz OnlineRecognizer.from_paraformersV{{3F#G$G$7  )  )(#+#  -%# )'A'A'A  3#%+5+&$"&  &&78'  rrZ ctc_graphctc_max_activecd|j|}t|t|t|}t| |}t |||||}t ||}t ||| }t| | }t|||||| ||t||| }t||_ ||_ |S) a Please refer to ``_ 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. rule_fsts: If not empty, it specifies fsts for inverse text normalization. If there are multiple fsts, they are separated by a comma. rule_fars: If not empty, it specifies fst archives for inverse text normalization. If there are multiple archives, they are separated by a comma. device: onnxruntime cuda device index. rZrr)zipformer2_ctcr r$rVr=rtrX)graph max_activer]) r`rarcctc_fst_decoder_configrdr0r>r?re) rfrrrr rrr rr rirjrk)rlr rZr$r%r&r,r-r.r/r0rxryr@r=r>r?rArMrNrOrmzipformer2_ctc_configrVrar`rcrros rfrom_zipformer2_ctcz$OnlineRecognizer.from_zipformer2_ctcs`{{3F#E" >U K(  )0#+  -%# )'A'A'A  ";%"  3#%+#95+&$"&  &&78'  rcH|j|}t|t|t|}t| |}t ||||| }t ||}t ||| }t||||| | |t|||}t||_ ||_ |S){ Please refer to ``_ to download pre-trained models. 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. provider: onnxruntime execution providers. Valid values are: cpu, cuda, coreml. debug: True to show meta data in the model. rule_fsts: If not empty, it specifies fsts for inverse text normalization. If there are multiple fsts, they are separated by a comma. rule_fars: If not empty, it specifies fst archives for inverse text normalization. If there are multiple archives, they are separated by a comma. device: onnxruntime cuda device index. r{rr) t_one_ctcr r$rVr=rtrXr]ru) rfrrrr rrrr rirjrk)rlr rZr$r%r&r,r-r.r/r0r@r=r>r?rArMrNrOrmt_one_ctc_configrVrar`rcros rfrom_t_one_ctczOnlineRecognizer.from_t_one_ctcsR{{3F#E"3 )  )&#+  -%# )'A'A'A  3#%+5+&$"&  &&78'  rcH|j|}t|t|t|}t| |}t ||||| }t ||}t ||| }t||||| | |t|||}t||_ ||_ |S)rr{rr)nemo_ctcr r$rVr=rtrXr]ru) rfrr rr rrrr rirjrk)rlr rZr$r%r&r,r-r.r/r0r@r=r>r?rArMrNrOrmnemo_ctc_configrVrar`rcros r from_nemo_ctczOnlineRecognizer.from_nemo_ctc sR{{3F#E"2 )  )$#+  -%# )'A'A'A  3#%+5+&$"&  &&78'  r chunk_sizenum_left_chunkscL|j|}t|t|t|||}t| |}t |||||}t ||}t | | | }t||||| ||t|||}t||_ ||_ |S)a Please refer to ``_ to download pre-trained models for different languages, e.g., Chinese, English, etc. Args: tokens: Path to ``tokens.txt``. Each line in ``tokens.txt`` contains two columns:: symbol integer_id model: Path to ``model.onnx``. chunk_size: The --chunk-size parameter from WeNet. num_left_chunks: The --num-left-chunks parameter from WeNet. num_threads: Number of threads for neural network computation. sample_rate: Sample rate of the training data used to train the model. feature_dim: Dimension of the feature used to train the model. enable_endpoint_detection: True to enable endpoint detection. False to disable endpoint detection. rule1_min_trailing_silence: Used only when enable_endpoint_detection is True. If the duration of trailing silence in seconds is larger than this value, we assume an endpoint is detected. rule2_min_trailing_silence: Used only when enable_endpoint_detection is True. If we have decoded something that is nonsilence and if the duration of trailing silence in seconds is larger than this value, we assume an endpoint is detected. rule3_min_utterance_length: Used only when enable_endpoint_detection is True. If the utterance length in seconds is larger than this value, we assume an endpoint is detected. decoding_method: The only valid value is greedy_search. provider: onnxruntime execution providers. Valid values are: cpu, cuda, coreml. rule_fsts: If not empty, it specifies fsts for inverse text normalization. If there are multiple fsts, they are separated by a comma. rule_fars: If not empty, it specifies fst archives for inverse text normalization. If there are multiple archives, they are separated by a comma. device: onnxruntime cuda device index. )rZrrrr) wenet_ctcr r$rVr=rtrXr]ru) rfrrrr rrrr rirjrk)rlr rZrrr$r%r&r,r-r.r/r0r@r=r>r?rArMrNrOrmwenet_ctc_configrVrar`rcros rfrom_wenet_ctczOnlineRecognizer.from_wenet_ctcs\{{3F#E"4!+  )  )&#+  -%# )'A'A'A  3#%+5+&$"&  &&78'  rNhotwordscp||jjS|jj|S)aSCreate a new online stream for streaming recognition. Args: hotwords: Optional hotwords string. Each word/phrase is separated by a space, and the tokens within a phrase are separated by the modeling unit separator. Returns: A new ``OnlineStream`` instance. )rj create_stream)rmrs rrzOnlineRecognizer.create_stream s2  ??002 2??00: :rsc:|jj|y)zzRun one decoding step on a single stream. Args: s: The ``OnlineStream`` to decode. N)rj decode_streamrmrs rrzOnlineRecognizer.decode_streams %%a(rssc:|jj|y)zRun one decoding step on multiple streams in parallel. Args: ss: A list of ``OnlineStream`` instances to decode. N)rjdecode_streams)rmrs rrzOnlineRecognizer.decode_streams$s &&r*rreturnc8|jj|S)zCheck whether the stream has enough data for decoding. Args: s: The ``OnlineStream`` to check. Returns: True if there is enough data to run one decoding step. )rjis_readyrs rrzOnlineRecognizer.is_ready-s''**rc8|jj|S)zGet the full recognition result object for a stream. Args: s: The ``OnlineStream`` to query. Returns: An ``OnlineRecognizerResult`` containing text, tokens, timestamps, and other fields. )rj get_resultrs rget_result_allzOnlineRecognizer.get_result_all9s))!,,rch|jj|jjS)zGet the recognition result as a plain string. Args: s: The ``OnlineStream`` to query. Returns: The recognized text with leading/trailing whitespace removed. )rjrtextstriprs rrzOnlineRecognizer.get_resultFs())!,117799rcT|jj|jS)zGet the recognition result as a JSON string. Args: s: The ``OnlineStream`` to query. Returns: A JSON-encoded string representing the recognition result. )rjras_json_stringrs rget_result_as_json_stringz*OnlineRecognizer.get_result_as_json_stringRs"))!,;;==rcL|jj|jS)zGet the list of recognized tokens for a stream. Args: s: The ``OnlineStream`` to query. Returns: A list of token strings from the recognition result. )rjrr rs rr zOnlineRecognizer.tokens^s))!,333rcL|jj|jS)a Get the timestamps of each recognized token. Args: s: The ``OnlineStream`` to query. Returns: A list of floats where each element is the time in seconds of the corresponding token from :meth:`tokens`. )rjr timestampsrs rrzOnlineRecognizer.timestampsjs))!,777rcL|jj|jS)zGet the start time of the current segment. Args: s: The ``OnlineStream`` to query. Returns: The start time in seconds for the current recognition segment. )rjr start_timers rrzOnlineRecognizer.start_timews))!,777rcL|jj|jS)zGet the output log probabilities for each recognized token. Args: s: The ``OnlineStream`` to query. Returns: A list of log-probability values, one per token from :meth:`tokens`. )rjrys_probsrs rrzOnlineRecognizer.ys_probs))!,555rcL|jj|jS)aGet the language model log probabilities for each recognized token. Args: s: The ``OnlineStream`` to query. Returns: A list of LM log-probability values, one per token from :meth:`tokens`. )rjrlm_probsrs rrzOnlineRecognizer.lm_probsrrcL|jj|jS)zGet the context biasing scores for each recognized token. Args: s: The ``OnlineStream`` to query. Returns: A list of context scores, one per token from :meth:`tokens`. )rjrcontext_scoresrs rrzOnlineRecognizer.context_scoress))!,;;;rc8|jj|S)zCheck whether an endpoint has been detected for the stream. Args: s: The ``OnlineStream`` to check. Returns: True if an endpoint (e.g. trailing silence) is detected. )rj is_endpointrs rrzOnlineRecognizer.is_endpoints**1--rc8|jj|S)a<Reset the stream state after an endpoint is detected. Call this after :meth:`is_endpoint` returns True to prepare the stream for the next utterance. Args: s: The ``OnlineStream`` to reset. Returns: True if the stream was successfully reset. )rjresetrs rrzOnlineRecognizer.resets$$Q''r).>P4@gyTFF333333@333333?r greedy_searchg?rrcjkcharrrg?Tg@FFrrcpuri TFTTrrFrrrrr)rrrFrrrrrFrrrrrr)rrrFrrrrri rFrrrrrr)ri@rFrrrrrFrrrrrr)rrrrFrrrrrFrrrrrr)N)"__name__ __module__ __qualname____doc__ classmethodstrintfloatboolrprwrrrrrrrrrrrrrrrr rrrrrrrrrrr"s <|"!"& */,/,/,0. ! #"&"&#&#&'&0,.%& $',(,(,%'%'#(gccc c  c  cccccc cc$(c%*c %*!c"%*#c$%c&'c()c*+c,-c./c01c23c4 5c67c8 9c:!;c<=c>?c@AcBCcDEcFGcH!$IcJ!$KcL'*McN #OcPQcR!%ScT"&UcV"&WcX #YcZ #[c\!]c^_c`acbccdecfgccJ  "*/,/,/,0.)~~~ ~  ~  ~~$(~%*~%*~%*~~~~~ !~"#~$%~&'~()~~@ "*/,/,/,0."+DDD D  D  D$(D%*D%*D%*DDDDDD !D"#D$%D&'D()D*+DDL !*/,/,/,0.'yyy y  y  y$(y%*y%*y%*yyyyyy !y"#y$%y&'yyv "*/,/,/,0.'yyy y  y  y$(y%*y%*y%*yyyyyy !y"#y$%y&'yyv  "*/,/,/,0.+@@@ @  @  @@@$(@%*@%*@%*@@@@ !@"#@$%@&'@()@*+@@D;hsm;")|)+l!3+ +, +4 + - -1G - :L :S : >< >C > 4 4c 4 8L 8T%[ 8 8L 8U 8 6, 64; 6 6, 64; 6 < <e < .\ .d . (| ( (rrN)pathlibrtypingrrsherpa_onnx.lib._sherpa_onnxrrrr r r r r rrrirrrrrrrrrrrobjectrrrrsI!   I   434`(v`(r