From 21e4310b5cbdb5ded4d86050c31654fc8a7c521f Mon Sep 17 00:00:00 2001 From: Varshith B Date: Sun, 29 Dec 2024 16:45:45 +0000 Subject: [PATCH] feat: streaming whisper --- audio_example.py | 32 +- nodes/audio_utils/__init__.py | 5 +- nodes/audio_utils/apply_whisper.py | 62 -- .../{save_asr_response.py => save_result.py} | 10 +- nodes/whisper_utils/__init__.py | 5 + nodes/whisper_utils/apply_whisper.py | 37 + nodes/whisper_utils/silero_vad_iterator.py | 146 +++ nodes/whisper_utils/whisper_online.py | 852 ++++++++++++++++++ src/comfystream/utils.py | 2 +- workflows/audio-whsiper-example-workflow.json | 6 +- 10 files changed, 1075 insertions(+), 82 deletions(-) delete mode 100644 nodes/audio_utils/apply_whisper.py rename nodes/audio_utils/{save_asr_response.py => save_result.py} (72%) create mode 100644 nodes/whisper_utils/__init__.py create mode 100644 nodes/whisper_utils/apply_whisper.py create mode 100644 nodes/whisper_utils/silero_vad_iterator.py create mode 100644 nodes/whisper_utils/whisper_online.py diff --git a/audio_example.py b/audio_example.py index 9b48cc6c..1d73a383 100644 --- a/audio_example.py +++ b/audio_example.py @@ -5,20 +5,36 @@ from comfystream.client import ComfyStreamClient async def main(): - cwd = "/home/user/ComfyUI" - client = ComfyStreamClient(cwd=cwd) - + cwd = "/home/user/ComfyUI" + client = ComfyStreamClient(cwd=cwd, type="audio") with open("./workflows/audio-whsiper-example-workflow.json", "r") as f: prompt = json.load(f) client.set_prompt(prompt) - - waveform, _ = torchaudio.load("/home/user/harvard.wav") + waveform, sr = torchaudio.load("/home/user/harvard.wav") + resampler = torchaudio.transforms.Resample(orig_freq=sr, new_freq=16000) + waveform = resampler(waveform) + sr = 16000 if waveform.ndim > 1: - audio_tensor = waveform.mean(dim=0) + waveform = waveform.mean(dim=0, keepdim=True) + + chunk_ms = 20 + chunk_size = int(sr * (chunk_ms / 1000.0)) + + total_samples = waveform.shape[1] + offset = 0 + + results = [] + while offset < total_samples: + end = min(offset + chunk_size, total_samples) + chunk = waveform[:, offset:end] + offset = end + results.append(await client.queue_prompt(chunk.numpy().squeeze())) - output = await client.queue_prompt(audio_tensor) - print(output) + print("Result:") + for result in results: + if result[0] is not None: + print(result[-1]) if __name__ == "__main__": asyncio.run(main()) \ No newline at end of file diff --git a/nodes/audio_utils/__init__.py b/nodes/audio_utils/__init__.py index 162e8244..b859b5c2 100644 --- a/nodes/audio_utils/__init__.py +++ b/nodes/audio_utils/__init__.py @@ -1,8 +1,7 @@ -from .apply_whisper import ApplyWhisper from .load_audio_tensor import LoadAudioTensor -from .save_asr_response import SaveASRResponse +from .save_result import SaveResult from .save_audio_tensor import SaveAudioTensor -NODE_CLASS_MAPPINGS = {"LoadAudioTensor": LoadAudioTensor, "SaveASRResponse": SaveASRResponse, "ApplyWhisper": ApplyWhisper, "SaveAudioTensor": SaveAudioTensor} +NODE_CLASS_MAPPINGS = {"LoadAudioTensor": LoadAudioTensor, "SaveResult": SaveResult, "SaveAudioTensor": SaveAudioTensor} __all__ = ["NODE_CLASS_MAPPINGS"] diff --git a/nodes/audio_utils/apply_whisper.py b/nodes/audio_utils/apply_whisper.py deleted file mode 100644 index c10ac884..00000000 --- a/nodes/audio_utils/apply_whisper.py +++ /dev/null @@ -1,62 +0,0 @@ -import torch -import whisper - -class ApplyWhisper: - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "audio": ("AUDIO",), - "model": (["base", "tiny", "small", "medium", "large"],), - } - } - - CATEGORY = "audio_utils" - RETURN_TYPES = ("DICT",) - FUNCTION = "apply_whisper" - - def __init__(self): - self.model = None - self.audio_buffer = [] - # TO:DO to get them as params - self.sample_rate = 16000 - self.min_duration = 1.0 - self.device = "cuda" if torch.cuda.is_available() else "cpu" - - def apply_whisper(self, audio, model): - if self.model is None: - self.model = whisper.load_model(model).to(self.device) - - self.audio_buffer.append(audio) - total_duration = sum(chunk.shape[0] / self.sample_rate for chunk in self.audio_buffer) - if total_duration < self.min_duration: - return {"text": "", "segments_alignment": [], "words_alignment": []} - - concatenated_audio = torch.cat(self.audio_buffer, dim=0).to(self.device) - self.audio_buffer = [] - result = self.model.transcribe(concatenated_audio.float(), fp16=True, word_timestamps=True) - segments = result["segments"] - segments_alignment = [] - words_alignment = [] - - for segment in segments: - segment_dict = { - "value": segment["text"].strip(), - "start": segment["start"], - "end": segment["end"] - } - segments_alignment.append(segment_dict) - - for word in segment["words"]: - word_dict = { - "value": word["word"].strip(), - "start": word["start"], - "end": word["end"] - } - words_alignment.append(word_dict) - - return ({ - "text": result["text"].strip(), - "segments_alignment": segments_alignment, - "words_alignment": words_alignment - },) diff --git a/nodes/audio_utils/save_asr_response.py b/nodes/audio_utils/save_result.py similarity index 72% rename from nodes/audio_utils/save_asr_response.py rename to nodes/audio_utils/save_result.py index 816ca1bc..4b8ef892 100644 --- a/nodes/audio_utils/save_asr_response.py +++ b/nodes/audio_utils/save_result.py @@ -1,6 +1,6 @@ from comfystream import tensor_cache -class SaveASRResponse: +class SaveResult: CATEGORY = "audio_utils" RETURN_TYPES = () FUNCTION = "execute" @@ -10,7 +10,7 @@ class SaveASRResponse: def INPUT_TYPES(s): return { "required": { - "data": ("DICT",), + "result": ("RESULT",), } } @@ -18,7 +18,7 @@ def INPUT_TYPES(s): def IS_CHANGED(s): return float("nan") - def execute(self, data: dict): + def execute(self, result): fut = tensor_cache.audio_outputs.pop() - fut.set_result(data) - return data \ No newline at end of file + fut.set_result(result) + return result \ No newline at end of file diff --git a/nodes/whisper_utils/__init__.py b/nodes/whisper_utils/__init__.py new file mode 100644 index 00000000..5aa83c0c --- /dev/null +++ b/nodes/whisper_utils/__init__.py @@ -0,0 +1,5 @@ +from .apply_whisper import ApplyWhisper + +NODE_CLASS_MAPPINGS = {"ApplyWhisper": ApplyWhisper} + +__all__ = ["NODE_CLASS_MAPPINGS"] diff --git a/nodes/whisper_utils/apply_whisper.py b/nodes/whisper_utils/apply_whisper.py new file mode 100644 index 00000000..291c1f8b --- /dev/null +++ b/nodes/whisper_utils/apply_whisper.py @@ -0,0 +1,37 @@ +from .whisper_online import FasterWhisperASR, VACOnlineASRProcessor + +class ApplyWhisper: + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "audio": ("AUDIO",), + } + } + + CATEGORY = "whisper_utils" + RETURN_TYPES = ("RESULT",) + FUNCTION = "apply_whisper" + + def __init__(self): + self.asr = FasterWhisperASR( + lan="en", + modelsize="large-v3", + cache_dir=None, + model_dir=None, + logfile=None + ) + self.asr.use_vad() + + self.online = VACOnlineASRProcessor( + online_chunk_size=0.5, + asr=self.asr, + tokenizer=None, + logfile=None, + buffer_trimming=("segment", 15) + ) + + def apply_whisper(self, audio): + self.online.insert_audio_chunk(audio) + result = self.online.process_iter() + return (result,) diff --git a/nodes/whisper_utils/silero_vad_iterator.py b/nodes/whisper_utils/silero_vad_iterator.py new file mode 100644 index 00000000..1eea7af3 --- /dev/null +++ b/nodes/whisper_utils/silero_vad_iterator.py @@ -0,0 +1,146 @@ +import torch + +# This is copied from silero-vad's vad_utils.py: +# https://github.com/snakers4/silero-vad/blob/f6b1294cb27590fb2452899df98fb234dfef1134/utils_vad.py#L340 +# (except changed defaults) + +# Their licence is MIT, same as ours: https://github.com/snakers4/silero-vad/blob/f6b1294cb27590fb2452899df98fb234dfef1134/LICENSE + +class VADIterator: + def __init__(self, + model, + threshold: float = 0.5, + sampling_rate: int = 16000, + min_silence_duration_ms: int = 500, # makes sense on one recording that I checked + speech_pad_ms: int = 100 # same + ): + + """ + Class for stream imitation + + Parameters + ---------- + model: preloaded .jit silero VAD model + + threshold: float (default - 0.5) + Speech threshold. Silero VAD outputs speech probabilities for each audio chunk, probabilities ABOVE this value are considered as SPEECH. + It is better to tune this parameter for each dataset separately, but "lazy" 0.5 is pretty good for most datasets. + + sampling_rate: int (default - 16000) + Currently silero VAD models support 8000 and 16000 sample rates + + min_silence_duration_ms: int (default - 100 milliseconds) + In the end of each speech chunk wait for min_silence_duration_ms before separating it + + speech_pad_ms: int (default - 30 milliseconds) + Final speech chunks are padded by speech_pad_ms each side + """ + + self.model = model + self.threshold = threshold + self.sampling_rate = sampling_rate + + if sampling_rate not in [8000, 16000]: + raise ValueError('VADIterator does not support sampling rates other than [8000, 16000]') + + self.min_silence_samples = sampling_rate * min_silence_duration_ms / 1000 + self.speech_pad_samples = sampling_rate * speech_pad_ms / 1000 + self.reset_states() + + def reset_states(self): + + self.model.reset_states() + self.triggered = False + self.temp_end = 0 + self.current_sample = 0 + + def __call__(self, x, return_seconds=False): + """ + x: torch.Tensor + audio chunk (see examples in repo) + + return_seconds: bool (default - False) + whether return timestamps in seconds (default - samples) + """ + + if not torch.is_tensor(x): + try: + x = torch.Tensor(x) + except: + raise TypeError("Audio cannot be casted to tensor. Cast it manually") + + window_size_samples = len(x[0]) if x.dim() == 2 else len(x) + self.current_sample += window_size_samples + + speech_prob = self.model(x, self.sampling_rate).item() + + if (speech_prob >= self.threshold) and self.temp_end: + self.temp_end = 0 + + if (speech_prob >= self.threshold) and not self.triggered: + self.triggered = True + speech_start = self.current_sample - self.speech_pad_samples + return {'start': int(speech_start) if not return_seconds else round(speech_start / self.sampling_rate, 1)} + + if (speech_prob < self.threshold - 0.15) and self.triggered: + if not self.temp_end: + self.temp_end = self.current_sample + if self.current_sample - self.temp_end < self.min_silence_samples: + return None + else: + speech_end = self.temp_end + self.speech_pad_samples + self.temp_end = 0 + self.triggered = False + return {'end': int(speech_end) if not return_seconds else round(speech_end / self.sampling_rate, 1)} + + return None + +####################### +# because Silero now requires exactly 512-sized audio chunks + +import numpy as np +class FixedVADIterator(VADIterator): + '''It fixes VADIterator by allowing to process any audio length, not only exactly 512 frames at once. + If audio to be processed at once is long and multiple voiced segments detected, + then __call__ returns the start of the first segment, and end (or middle, which means no end) of the last segment. + ''' + + def reset_states(self): + super().reset_states() + self.buffer = np.array([],dtype=np.float32) + + def __call__(self, x, return_seconds=False): + self.buffer = np.append(self.buffer, x) + ret = None + while len(self.buffer) >= 512: + r = super().__call__(self.buffer[:512], return_seconds=return_seconds) + self.buffer = self.buffer[512:] + if ret is None: + ret = r + elif r is not None: + if 'end' in r: + ret['end'] = r['end'] # the latter end + if 'start' in r and 'end' in ret: # there is an earlier start. + # Remove end, merging this segment with the previous one. + del ret['end'] + return ret if ret != {} else None + +if __name__ == "__main__": + # test/demonstrate the need for FixedVADIterator: + + import torch + model, _ = torch.hub.load( + repo_or_dir='snakers4/silero-vad', + model='silero_vad' + ) + vac = FixedVADIterator(model) +# vac = VADIterator(model) # the second case crashes with this + + # this works: for both + audio_buffer = np.array([0]*(512),dtype=np.float32) + vac(audio_buffer) + + # this crashes on the non FixedVADIterator with + # ops.prim.RaiseException("Input audio chunk is too short", "builtins.ValueError") + audio_buffer = np.array([0]*(512-1),dtype=np.float32) + vac(audio_buffer) diff --git a/nodes/whisper_utils/whisper_online.py b/nodes/whisper_utils/whisper_online.py new file mode 100644 index 00000000..7d6cf763 --- /dev/null +++ b/nodes/whisper_utils/whisper_online.py @@ -0,0 +1,852 @@ +#!/usr/bin/env python3 +import sys +import numpy as np +import librosa +from functools import lru_cache +import time +import logging + +import io +import math + +logger = logging.getLogger(__name__) + +@lru_cache(10**6) +def load_audio(fname): + a, _ = librosa.load(fname, sr=16000, dtype=np.float32) + return a + +def load_audio_chunk(fname, beg, end): + audio = load_audio(fname) + beg_s = int(beg*16000) + end_s = int(end*16000) + return audio[beg_s:end_s] + + +# Whisper backend + +class ASRBase: + + sep = " " # join transcribe words with this character (" " for whisper_timestamped, + # "" for faster-whisper because it emits the spaces when neeeded) + + def __init__(self, lan, modelsize=None, cache_dir=None, model_dir=None, logfile=sys.stderr): + self.logfile = logfile + + self.transcribe_kargs = {} + if lan == "auto": + self.original_language = None + else: + self.original_language = lan + + self.model = self.load_model(modelsize, cache_dir, model_dir) + + + def load_model(self, modelsize, cache_dir): + raise NotImplemented("must be implemented in the child class") + + def transcribe(self, audio, init_prompt=""): + raise NotImplemented("must be implemented in the child class") + + def use_vad(self): + raise NotImplemented("must be implemented in the child class") + + +class WhisperTimestampedASR(ASRBase): + """Uses whisper_timestamped library as the backend. Initially, we tested the code on this backend. It worked, but slower than faster-whisper. + On the other hand, the installation for GPU could be easier. + """ + + sep = " " + + def load_model(self, modelsize=None, cache_dir=None, model_dir=None): + import whisper + import whisper_timestamped + from whisper_timestamped import transcribe_timestamped + self.transcribe_timestamped = transcribe_timestamped + if model_dir is not None: + logger.debug("ignoring model_dir, not implemented") + return whisper.load_model(modelsize, download_root=cache_dir) + + def transcribe(self, audio, init_prompt=""): + result = self.transcribe_timestamped(self.model, + audio, language=self.original_language, + initial_prompt=init_prompt, verbose=None, + condition_on_previous_text=True, **self.transcribe_kargs) + return result + + def ts_words(self,r): + # return: transcribe result object to [(beg,end,"word1"), ...] + o = [] + for s in r["segments"]: + for w in s["words"]: + t = (w["start"],w["end"],w["text"]) + o.append(t) + return o + + def segments_end_ts(self, res): + return [s["end"] for s in res["segments"]] + + def use_vad(self): + self.transcribe_kargs["vad"] = True + + def set_translate_task(self): + self.transcribe_kargs["task"] = "translate" + + + + +class FasterWhisperASR(ASRBase): + """Uses faster-whisper library as the backend. Works much faster, appx 4-times (in offline mode). For GPU, it requires installation with a specific CUDNN version. + """ + + sep = "" + + def load_model(self, modelsize=None, cache_dir=None, model_dir=None): + from faster_whisper import WhisperModel +# logging.getLogger("faster_whisper").setLevel(logger.level) + if model_dir is not None: + logger.debug(f"Loading whisper model from model_dir {model_dir}. modelsize and cache_dir parameters are not used.") + model_size_or_path = model_dir + elif modelsize is not None: + model_size_or_path = modelsize + else: + raise ValueError("modelsize or model_dir parameter must be set") + + + # this worked fast and reliably on NVIDIA L40 + model = WhisperModel(model_size_or_path, device="cuda", compute_type="float16", download_root=cache_dir) + + # or run on GPU with INT8 + # tested: the transcripts were different, probably worse than with FP16, and it was slightly (appx 20%) slower + #model = WhisperModel(model_size, device="cuda", compute_type="int8_float16") + + # or run on CPU with INT8 + # tested: works, but slow, appx 10-times than cuda FP16 +# model = WhisperModel(modelsize, device="cpu", compute_type="int8") #, download_root="faster-disk-cache-dir/") + return model + + def transcribe(self, audio, init_prompt=""): + + # tested: beam_size=5 is faster and better than 1 (on one 200 second document from En ESIC, min chunk 0.01) + segments, info = self.model.transcribe(audio, language=self.original_language, initial_prompt=init_prompt, beam_size=5, word_timestamps=True, condition_on_previous_text=True, **self.transcribe_kargs) + #print(info) # info contains language detection result + + return list(segments) + + def ts_words(self, segments): + o = [] + for segment in segments: + for word in segment.words: + if segment.no_speech_prob > 0.9: + continue + # not stripping the spaces -- should not be merged with them! + w = word.word + t = (word.start, word.end, w) + o.append(t) + return o + + def segments_end_ts(self, res): + return [s.end for s in res] + + def use_vad(self): + self.transcribe_kargs["vad_filter"] = True + + def set_translate_task(self): + self.transcribe_kargs["task"] = "translate" + + +class OpenaiApiASR(ASRBase): + """Uses OpenAI's Whisper API for audio transcription.""" + + def __init__(self, lan=None, temperature=0, logfile=sys.stderr): + self.logfile = logfile + + self.modelname = "whisper-1" + self.original_language = None if lan == "auto" else lan # ISO-639-1 language code + self.response_format = "verbose_json" + self.temperature = temperature + + self.load_model() + + self.use_vad_opt = False + + # reset the task in set_translate_task + self.task = "transcribe" + + def load_model(self, *args, **kwargs): + from openai import OpenAI + self.client = OpenAI() + + self.transcribed_seconds = 0 # for logging how many seconds were processed by API, to know the cost + + + def ts_words(self, segments): + no_speech_segments = [] + if self.use_vad_opt: + for segment in segments.segments: + # TODO: threshold can be set from outside + if segment["no_speech_prob"] > 0.8: + no_speech_segments.append((segment.get("start"), segment.get("end"))) + + o = [] + for word in segments.words: + start = word.start + end = word.end + if any(s[0] <= start <= s[1] for s in no_speech_segments): + # print("Skipping word", word.get("word"), "because it's in a no-speech segment") + continue + o.append((start, end, word.word)) + return o + + + def segments_end_ts(self, res): + return [s.end for s in res.words] + + def transcribe(self, audio_data, prompt=None, *args, **kwargs): + # Write the audio data to a buffer + buffer = io.BytesIO() + buffer.name = "temp.wav" + sf.write(buffer, audio_data, samplerate=16000, format='WAV', subtype='PCM_16') + buffer.seek(0) # Reset buffer's position to the beginning + + self.transcribed_seconds += math.ceil(len(audio_data)/16000) # it rounds up to the whole seconds + + params = { + "model": self.modelname, + "file": buffer, + "response_format": self.response_format, + "temperature": self.temperature, + "timestamp_granularities": ["word", "segment"] + } + if self.task != "translate" and self.original_language: + params["language"] = self.original_language + if prompt: + params["prompt"] = prompt + + if self.task == "translate": + proc = self.client.audio.translations + else: + proc = self.client.audio.transcriptions + + # Process transcription/translation + transcript = proc.create(**params) + logger.debug(f"OpenAI API processed accumulated {self.transcribed_seconds} seconds") + + return transcript + + def use_vad(self): + self.use_vad_opt = True + + def set_translate_task(self): + self.task = "translate" + + + + +class HypothesisBuffer: + + def __init__(self, logfile=sys.stderr): + self.commited_in_buffer = [] + self.buffer = [] + self.new = [] + + self.last_commited_time = 0 + self.last_commited_word = None + + self.logfile = logfile + + def insert(self, new, offset): + # compare self.commited_in_buffer and new. It inserts only the words in new that extend the commited_in_buffer, it means they are roughly behind last_commited_time and new in content + # the new tail is added to self.new + + new = [(a+offset,b+offset,t) for a,b,t in new] + self.new = [(a,b,t) for a,b,t in new if a > self.last_commited_time-0.1] + + if len(self.new) >= 1: + a,b,t = self.new[0] + if abs(a - self.last_commited_time) < 1: + if self.commited_in_buffer: + # it's going to search for 1, 2, ..., 5 consecutive words (n-grams) that are identical in commited and new. If they are, they're dropped. + cn = len(self.commited_in_buffer) + nn = len(self.new) + for i in range(1,min(min(cn,nn),5)+1): # 5 is the maximum + c = " ".join([self.commited_in_buffer[-j][2] for j in range(1,i+1)][::-1]) + tail = " ".join(self.new[j-1][2] for j in range(1,i+1)) + if c == tail: + words = [] + for j in range(i): + words.append(repr(self.new.pop(0))) + words_msg = " ".join(words) + logger.debug(f"removing last {i} words: {words_msg}") + break + + def flush(self): + # returns commited chunk = the longest common prefix of 2 last inserts. + + commit = [] + while self.new: + na, nb, nt = self.new[0] + + if len(self.buffer) == 0: + break + + if nt == self.buffer[0][2]: + commit.append((na,nb,nt)) + self.last_commited_word = nt + self.last_commited_time = nb + self.buffer.pop(0) + self.new.pop(0) + else: + break + self.buffer = self.new + self.new = [] + self.commited_in_buffer.extend(commit) + return commit + + def pop_commited(self, time): + while self.commited_in_buffer and self.commited_in_buffer[0][1] <= time: + self.commited_in_buffer.pop(0) + + def complete(self): + return self.buffer + +class OnlineASRProcessor: + + SAMPLING_RATE = 16000 + + def __init__(self, asr, tokenizer=None, buffer_trimming=("segment", 15), logfile=sys.stderr): + """asr: WhisperASR object + tokenizer: sentence tokenizer object for the target language. Must have a method *split* that behaves like the one of MosesTokenizer. It can be None, if "segment" buffer trimming option is used, then tokenizer is not used at all. + ("segment", 15) + buffer_trimming: a pair of (option, seconds), where option is either "sentence" or "segment", and seconds is a number. Buffer is trimmed if it is longer than "seconds" threshold. Default is the most recommended option. + logfile: where to store the log. + """ + self.asr = asr + self.tokenizer = tokenizer + self.logfile = logfile + + self.init() + + self.buffer_trimming_way, self.buffer_trimming_sec = buffer_trimming + + def init(self, offset=None): + """run this when starting or restarting processing""" + self.audio_buffer = np.array([],dtype=np.float32) + self.transcript_buffer = HypothesisBuffer(logfile=self.logfile) + self.buffer_time_offset = 0 + if offset is not None: + self.buffer_time_offset = offset + self.transcript_buffer.last_commited_time = self.buffer_time_offset + self.commited = [] + + def insert_audio_chunk(self, audio): + self.audio_buffer = np.append(self.audio_buffer, audio) + + def prompt(self): + """Returns a tuple: (prompt, context), where "prompt" is a 200-character suffix of commited text that is inside of the scrolled away part of audio buffer. + "context" is the commited text that is inside the audio buffer. It is transcribed again and skipped. It is returned only for debugging and logging reasons. + """ + k = max(0,len(self.commited)-1) + while k > 0 and self.commited[k-1][1] > self.buffer_time_offset: + k -= 1 + + p = self.commited[:k] + p = [t for _,_,t in p] + prompt = [] + l = 0 + while p and l < 200: # 200 characters prompt size + x = p.pop(-1) + l += len(x)+1 + prompt.append(x) + non_prompt = self.commited[k:] + return self.asr.sep.join(prompt[::-1]), self.asr.sep.join(t for _,_,t in non_prompt) + + def process_iter(self): + """Runs on the current audio buffer. + Returns: a tuple (beg_timestamp, end_timestamp, "text"), or (None, None, ""). + The non-emty text is confirmed (committed) partial transcript. + """ + + prompt, non_prompt = self.prompt() + logger.debug(f"PROMPT: {prompt}") + logger.debug(f"CONTEXT: {non_prompt}") + logger.debug(f"transcribing {len(self.audio_buffer)/self.SAMPLING_RATE:2.2f} seconds from {self.buffer_time_offset:2.2f}") + res = self.asr.transcribe(self.audio_buffer, init_prompt=prompt) + + # transform to [(beg,end,"word1"), ...] + tsw = self.asr.ts_words(res) + + self.transcript_buffer.insert(tsw, self.buffer_time_offset) + o = self.transcript_buffer.flush() + self.commited.extend(o) + completed = self.to_flush(o) + logger.debug(f">>>>COMPLETE NOW: {completed}") + the_rest = self.to_flush(self.transcript_buffer.complete()) + logger.debug(f"INCOMPLETE: {the_rest}") + + # there is a newly confirmed text + + if o and self.buffer_trimming_way == "sentence": # trim the completed sentences + if len(self.audio_buffer)/self.SAMPLING_RATE > self.buffer_trimming_sec: # longer than this + self.chunk_completed_sentence() + + + if self.buffer_trimming_way == "segment": + s = self.buffer_trimming_sec # trim the completed segments longer than s, + else: + s = 30 # if the audio buffer is longer than 30s, trim it + + if len(self.audio_buffer)/self.SAMPLING_RATE > s: + self.chunk_completed_segment(res) + + # alternative: on any word + #l = self.buffer_time_offset + len(self.audio_buffer)/self.SAMPLING_RATE - 10 + # let's find commited word that is less + #k = len(self.commited)-1 + #while k>0 and self.commited[k][1] > l: + # k -= 1 + #t = self.commited[k][1] + logger.debug("chunking segment") + #self.chunk_at(t) + + logger.debug(f"len of buffer now: {len(self.audio_buffer)/self.SAMPLING_RATE:2.2f}") + return self.to_flush(o) + + def chunk_completed_sentence(self): + if self.commited == []: return + logger.debug(self.commited) + sents = self.words_to_sentences(self.commited) + for s in sents: + logger.debug(f"\t\tSENT: {s}") + if len(sents) < 2: + return + while len(sents) > 2: + sents.pop(0) + # we will continue with audio processing at this timestamp + chunk_at = sents[-2][1] + + logger.debug(f"--- sentence chunked at {chunk_at:2.2f}") + self.chunk_at(chunk_at) + + def chunk_completed_segment(self, res): + if self.commited == []: return + + ends = self.asr.segments_end_ts(res) + + t = self.commited[-1][1] + + if len(ends) > 1: + + e = ends[-2]+self.buffer_time_offset + while len(ends) > 2 and e > t: + ends.pop(-1) + e = ends[-2]+self.buffer_time_offset + if e <= t: + logger.debug(f"--- segment chunked at {e:2.2f}") + self.chunk_at(e) + else: + logger.debug(f"--- last segment not within commited area") + else: + logger.debug(f"--- not enough segments to chunk") + + + + + + def chunk_at(self, time): + """trims the hypothesis and audio buffer at "time" + """ + self.transcript_buffer.pop_commited(time) + cut_seconds = time - self.buffer_time_offset + self.audio_buffer = self.audio_buffer[int(cut_seconds*self.SAMPLING_RATE):] + self.buffer_time_offset = time + + def words_to_sentences(self, words): + """Uses self.tokenizer for sentence segmentation of words. + Returns: [(beg,end,"sentence 1"),...] + """ + + cwords = [w for w in words] + t = " ".join(o[2] for o in cwords) + s = self.tokenizer.split(t) + out = [] + while s: + beg = None + end = None + sent = s.pop(0).strip() + fsent = sent + while cwords: + b,e,w = cwords.pop(0) + w = w.strip() + if beg is None and sent.startswith(w): + beg = b + elif end is None and sent == w: + end = e + out.append((beg,end,fsent)) + break + sent = sent[len(w):].strip() + return out + + def finish(self): + """Flush the incomplete text when the whole processing ends. + Returns: the same format as self.process_iter() + """ + o = self.transcript_buffer.complete() + f = self.to_flush(o) + logger.debug(f"last, noncommited: {f}") + self.buffer_time_offset += len(self.audio_buffer)/16000 + return f + + + def to_flush(self, sents, sep=None, offset=0, ): + # concatenates the timestamped words or sentences into one sequence that is flushed in one line + # sents: [(beg1, end1, "sentence1"), ...] or [] if empty + # return: (beg1,end-of-last-sentence,"concatenation of sentences") or (None, None, "") if empty + if sep is None: + sep = self.asr.sep + t = sep.join(s[2] for s in sents) + if len(sents) == 0: + b = None + e = None + else: + b = offset + sents[0][0] + e = offset + sents[-1][1] + return (b,e,t) + +class VACOnlineASRProcessor(OnlineASRProcessor): + '''Wraps OnlineASRProcessor with VAC (Voice Activity Controller). + + It works the same way as OnlineASRProcessor: it receives chunks of audio (e.g. 0.04 seconds), + it runs VAD and continuously detects whether there is speech or not. + When it detects end of speech (non-voice for 500ms), it makes OnlineASRProcessor to end the utterance immediately. + ''' + + def __init__(self, online_chunk_size, *a, **kw): + self.online_chunk_size = online_chunk_size + + self.online = OnlineASRProcessor(*a, **kw) + + # VAC: + import torch + model, _ = torch.hub.load( + repo_or_dir='snakers4/silero-vad', + model='silero_vad' + ) + from .silero_vad_iterator import FixedVADIterator + self.vac = FixedVADIterator(model) # we use the default options there: 500ms silence, 100ms padding, etc. + + self.logfile = self.online.logfile + self.init() + + def init(self): + self.online.init() + self.vac.reset_states() + self.current_online_chunk_buffer_size = 0 + + self.is_currently_final = False + + self.status = None # or "voice" or "nonvoice" + self.audio_buffer = np.array([],dtype=np.float32) + self.buffer_offset = 0 # in frames + + def clear_buffer(self): + self.buffer_offset += len(self.audio_buffer) + self.audio_buffer = np.array([],dtype=np.float32) + + + def insert_audio_chunk(self, audio): + res = self.vac(audio) + self.audio_buffer = np.append(self.audio_buffer, audio) + + if res is not None: + frame = list(res.values())[0]-self.buffer_offset + if 'start' in res and 'end' not in res: + self.status = 'voice' + send_audio = self.audio_buffer[frame:] + self.online.init(offset=(frame+self.buffer_offset)/self.SAMPLING_RATE) + self.online.insert_audio_chunk(send_audio) + self.current_online_chunk_buffer_size += len(send_audio) + self.clear_buffer() + elif 'end' in res and 'start' not in res: + self.status = 'nonvoice' + send_audio = self.audio_buffer[:frame] + self.online.insert_audio_chunk(send_audio) + self.current_online_chunk_buffer_size += len(send_audio) + self.is_currently_final = True + self.clear_buffer() + else: + beg = res["start"]-self.buffer_offset + end = res["end"]-self.buffer_offset + self.status = 'nonvoice' + send_audio = self.audio_buffer[beg:end] + self.online.init(offset=(beg+self.buffer_offset)/self.SAMPLING_RATE) + self.online.insert_audio_chunk(send_audio) + self.current_online_chunk_buffer_size += len(send_audio) + self.is_currently_final = True + self.clear_buffer() + else: + if self.status == 'voice': + self.online.insert_audio_chunk(self.audio_buffer) + self.current_online_chunk_buffer_size += len(self.audio_buffer) + self.clear_buffer() + else: + # We keep 1 second because VAD may later find start of voice in it. + # But we trim it to prevent OOM. + self.buffer_offset += max(0,len(self.audio_buffer)-self.SAMPLING_RATE) + self.audio_buffer = self.audio_buffer[-self.SAMPLING_RATE:] + + + def process_iter(self): + if self.is_currently_final: + return self.finish() + elif self.current_online_chunk_buffer_size > self.SAMPLING_RATE*self.online_chunk_size: + self.current_online_chunk_buffer_size = 0 + ret = self.online.process_iter() + return ret + else: + print("no online update, only VAD", self.status, file=self.logfile) + return (None, None, "") + + def finish(self): + ret = self.online.finish() + self.current_online_chunk_buffer_size = 0 + self.is_currently_final = False + return ret + + + +WHISPER_LANG_CODES = "af,am,ar,as,az,ba,be,bg,bn,bo,br,bs,ca,cs,cy,da,de,el,en,es,et,eu,fa,fi,fo,fr,gl,gu,ha,haw,he,hi,hr,ht,hu,hy,id,is,it,ja,jw,ka,kk,km,kn,ko,la,lb,ln,lo,lt,lv,mg,mi,mk,ml,mn,mr,ms,mt,my,ne,nl,nn,no,oc,pa,pl,ps,pt,ro,ru,sa,sd,si,sk,sl,sn,so,sq,sr,su,sv,sw,ta,te,tg,th,tk,tl,tr,tt,uk,ur,uz,vi,yi,yo,zh".split(",") + +def create_tokenizer(lan): + """returns an object that has split function that works like the one of MosesTokenizer""" + + assert lan in WHISPER_LANG_CODES, "language must be Whisper's supported lang code: " + " ".join(WHISPER_LANG_CODES) + + if lan == "uk": + import tokenize_uk + class UkrainianTokenizer: + def split(self, text): + return tokenize_uk.tokenize_sents(text) + return UkrainianTokenizer() + + # supported by fast-mosestokenizer + if lan in "as bn ca cs de el en es et fi fr ga gu hi hu is it kn lt lv ml mni mr nl or pa pl pt ro ru sk sl sv ta te yue zh".split(): + from mosestokenizer import MosesTokenizer + return MosesTokenizer(lan) + + # the following languages are in Whisper, but not in wtpsplit: + if lan in "as ba bo br bs fo haw hr ht jw lb ln lo mi nn oc sa sd sn so su sw tk tl tt".split(): + logger.debug(f"{lan} code is not supported by wtpsplit. Going to use None lang_code option.") + lan = None + + from wtpsplit import WtP + # downloads the model from huggingface on the first use + wtp = WtP("wtp-canine-s-12l-no-adapters") + class WtPtok: + def split(self, sent): + return wtp.split(sent, lang_code=lan) + return WtPtok() + + +def add_shared_args(parser): + """shared args for simulation (this entry point) and server + parser: argparse.ArgumentParser object + """ + parser.add_argument('--min-chunk-size', type=float, default=1.0, help='Minimum audio chunk size in seconds. It waits up to this time to do processing. If the processing takes shorter time, it waits, otherwise it processes the whole segment that was received by this time.') + parser.add_argument('--model', type=str, default='large-v2', choices="tiny.en,tiny,base.en,base,small.en,small,medium.en,medium,large-v1,large-v2,large-v3,large,large-v3-turbo".split(","),help="Name size of the Whisper model to use (default: large-v2). The model is automatically downloaded from the model hub if not present in model cache dir.") + parser.add_argument('--model_cache_dir', type=str, default=None, help="Overriding the default model cache dir where models downloaded from the hub are saved") + parser.add_argument('--model_dir', type=str, default=None, help="Dir where Whisper model.bin and other files are saved. This option overrides --model and --model_cache_dir parameter.") + parser.add_argument('--lan', '--language', type=str, default='auto', help="Source language code, e.g. en,de,cs, or 'auto' for language detection.") + parser.add_argument('--task', type=str, default='transcribe', choices=["transcribe","translate"],help="Transcribe or translate.") + parser.add_argument('--backend', type=str, default="faster-whisper", choices=["faster-whisper", "whisper_timestamped", "openai-api"],help='Load only this backend for Whisper processing.') + parser.add_argument('--vac', action="store_true", default=False, help='Use VAC = voice activity controller. Recommended. Requires torch.') + parser.add_argument('--vac-chunk-size', type=float, default=0.04, help='VAC sample size in seconds.') + parser.add_argument('--vad', action="store_true", default=False, help='Use VAD = voice activity detection, with the default parameters.') + parser.add_argument('--buffer_trimming', type=str, default="segment", choices=["sentence", "segment"],help='Buffer trimming strategy -- trim completed sentences marked with punctuation mark and detected by sentence segmenter, or the completed segments returned by Whisper. Sentence segmenter must be installed for "sentence" option.') + parser.add_argument('--buffer_trimming_sec', type=float, default=15, help='Buffer trimming length threshold in seconds. If buffer length is longer, trimming sentence/segment is triggered.') + parser.add_argument("-l", "--log-level", dest="log_level", choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'], help="Set the log level", default='DEBUG') + +def asr_factory(args, logfile=sys.stderr): + """ + Creates and configures an ASR and ASR Online instance based on the specified backend and arguments. + """ + backend = args.backend + if backend == "openai-api": + logger.debug("Using OpenAI API.") + asr = OpenaiApiASR(lan=args.lan) + else: + if backend == "faster-whisper": + asr_cls = FasterWhisperASR + else: + asr_cls = WhisperTimestampedASR + + # Only for FasterWhisperASR and WhisperTimestampedASR + size = args.model + t = time.time() + logger.info(f"Loading Whisper {size} model for {args.lan}...") + asr = asr_cls(modelsize=size, lan=args.lan, cache_dir=args.model_cache_dir, model_dir=args.model_dir) + e = time.time() + logger.info(f"done. It took {round(e-t,2)} seconds.") + + # Apply common configurations + if getattr(args, 'vad', False): # Checks if VAD argument is present and True + logger.info("Setting VAD filter") + asr.use_vad() + + language = args.lan + if args.task == "translate": + asr.set_translate_task() + tgt_language = "en" # Whisper translates into English + else: + tgt_language = language # Whisper transcribes in this language + + # Create the tokenizer + if args.buffer_trimming == "sentence": + tokenizer = create_tokenizer(tgt_language) + else: + tokenizer = None + + # Create the OnlineASRProcessor + if args.vac: + + online = VACOnlineASRProcessor(args.min_chunk_size, asr,tokenizer,logfile=logfile,buffer_trimming=(args.buffer_trimming, args.buffer_trimming_sec)) + else: + online = OnlineASRProcessor(asr,tokenizer,logfile=logfile,buffer_trimming=(args.buffer_trimming, args.buffer_trimming_sec)) + + return asr, online + +def set_logging(args,logger,other="_server"): + logging.basicConfig(#format='%(name)s + format='%(levelname)s\t%(message)s') + logger.setLevel(args.log_level) + logging.getLogger("whisper_online"+other).setLevel(args.log_level) +# logging.getLogger("whisper_online_server").setLevel(args.log_level) + + + +if __name__ == "__main__": + + import argparse + parser = argparse.ArgumentParser() + parser.add_argument('audio_path', type=str, help="Filename of 16kHz mono channel wav, on which live streaming is simulated.") + add_shared_args(parser) + parser.add_argument('--start_at', type=float, default=0.0, help='Start processing audio at this time.') + parser.add_argument('--offline', action="store_true", default=False, help='Offline mode.') + parser.add_argument('--comp_unaware', action="store_true", default=False, help='Computationally unaware simulation.') + + args = parser.parse_args() + + # reset to store stderr to different file stream, e.g. open(os.devnull,"w") + logfile = sys.stderr + + if args.offline and args.comp_unaware: + logger.error("No or one option from --offline and --comp_unaware are available, not both. Exiting.") + sys.exit(1) + +# if args.log_level: +# logging.basicConfig(format='whisper-%(levelname)s:%(name)s: %(message)s', +# level=getattr(logging, args.log_level)) + + set_logging(args,logger) + + audio_path = args.audio_path + + SAMPLING_RATE = 16000 + duration = len(load_audio(audio_path))/SAMPLING_RATE + logger.info("Audio duration is: %2.2f seconds" % duration) + + asr, online = asr_factory(args, logfile=logfile) + if args.vac: + min_chunk = args.vac_chunk_size + else: + min_chunk = args.min_chunk_size + + # load the audio into the LRU cache before we start the timer + a = load_audio_chunk(audio_path,0,1) + + # warm up the ASR because the very first transcribe takes much more time than the other + asr.transcribe(a) + + beg = args.start_at + start = time.time()-beg + + def output_transcript(o, now=None): + # output format in stdout is like: + # 4186.3606 0 1720 Takhle to je + # - the first three words are: + # - emission time from beginning of processing, in milliseconds + # - beg and end timestamp of the text segment, as estimated by Whisper model. The timestamps are not accurate, but they're useful anyway + # - the next words: segment transcript + if now is None: + now = time.time()-start + if o[0] is not None: + print("%1.4f %1.0f %1.0f %s" % (now*1000, o[0]*1000,o[1]*1000,o[2]),file=logfile,flush=True) + print("%1.4f %1.0f %1.0f %s" % (now*1000, o[0]*1000,o[1]*1000,o[2]),flush=True) + else: + # No text, so no output + pass + + if args.offline: ## offline mode processing (for testing/debugging) + a = load_audio(audio_path) + online.insert_audio_chunk(a) + try: + o = online.process_iter() + except AssertionError as e: + logger.error(f"assertion error: {repr(e)}") + else: + output_transcript(o) + now = None + elif args.comp_unaware: # computational unaware mode + end = beg + min_chunk + while True: + a = load_audio_chunk(audio_path,beg,end) + online.insert_audio_chunk(a) + try: + o = online.process_iter() + except AssertionError as e: + logger.error(f"assertion error: {repr(e)}") + pass + else: + output_transcript(o, now=end) + + logger.debug(f"## last processed {end:.2f}s") + + if end >= duration: + break + + beg = end + + if end + min_chunk > duration: + end = duration + else: + end += min_chunk + now = duration + + else: # online = simultaneous mode + end = 0 + while True: + now = time.time() - start + if now < end+min_chunk: + time.sleep(min_chunk+end-now) + end = time.time() - start + a = load_audio_chunk(audio_path,beg,end) + beg = end + online.insert_audio_chunk(a) + + try: + o = online.process_iter() + except AssertionError as e: + logger.error(f"assertion error: {e}") + pass + else: + output_transcript(o) + now = time.time() - start + logger.debug(f"## last processed {end:.2f} s, now is {now:.2f}, the latency is {now-end:.2f}") + + if end >= duration: + break + now = None + + o = online.finish() + output_transcript(o, now=now) diff --git a/src/comfystream/utils.py b/src/comfystream/utils.py index 0948e2b9..c9514aa7 100644 --- a/src/comfystream/utils.py +++ b/src/comfystream/utils.py @@ -48,7 +48,7 @@ def convert_prompt(prompt: PromptDictInput) -> Prompt: num_primary_inputs += 1 elif class_type in ["LoadImage", "LoadTensor", "LoadAudioTensor"]: num_inputs += 1 - elif class_type in ["PreviewImage", "SaveImage", "SaveTensor", "SaveASRResponse", "SaveAudioTensor"]: + elif class_type in ["PreviewImage", "SaveImage", "SaveTensor", "SaveResult", "SaveAudioTensor"]: num_outputs += 1 # Only handle single primary input diff --git a/workflows/audio-whsiper-example-workflow.json b/workflows/audio-whsiper-example-workflow.json index 95084f77..8926c080 100644 --- a/workflows/audio-whsiper-example-workflow.json +++ b/workflows/audio-whsiper-example-workflow.json @@ -21,14 +21,14 @@ }, "3": { "inputs": { - "data": [ + "result": [ "2", 0 ] }, - "class_type": "SaveASRResponse", + "class_type": "SaveResult", "_meta": { - "title": "Save ASR Response" + "title": "Save Result" } } } \ No newline at end of file