#!/usr/bin/env python3 import sys import numpy as np import librosa from functools import lru_cache import time @lru_cache def load_audio(fname): a, _ = librosa.load(fname, sr=16000) 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: # join transcribe words with this character (" " for whisper_timestamped, "" for faster-whisper because it emits the spaces when neeeded) sep = " " def __init__(self, lan, modelsize=None, cache_dir=None, model_dir=None): self.transcribe_kargs = {} 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") ## requires imports: # import whisper # import whisper_timestamped 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. If used, requires imports: import whisper import whisper_timestamped """ def load_model(self, modelsize=None, cache_dir=None, model_dir=None): if model_dir is not None: print("ignoring model_dir, not implemented",file=sys.stderr) return whisper.load_model(modelsize, download_root=cache_dir) def transcribe(self, audio, init_prompt=""): result = whisper_timestamped.transcribe_timestamped(self.model, audio, language=self.original_language, initial_prompt=init_prompt, verbose=None, condition_on_previous_text=True) 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): raise NotImplemented("Feature use_vad is not implemented for whisper_timestamped backend.") 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. Requires imports, if used: import faster_whisper """ sep = "" def load_model(self, modelsize=None, cache_dir=None, model_dir=None): from faster_whisper import WhisperModel if model_dir is not None: print(f"Loading whisper model from model_dir {model_dir}. modelsize and cache_dir parameters are not used.",file=sys.stderr) 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) return list(segments) def ts_words(self, segments): o = [] for segment in segments: for word in segment.words: # 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 HypothesisBuffer: def __init__(self): self.commited_in_buffer = [] self.buffer = [] self.new = [] self.last_commited_time = 0 self.last_commited_word = None 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: print("removing last",i,"words:",file=sys.stderr) for j in range(i): print("\t",self.new.pop(0),file=sys.stderr) 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): """asr: WhisperASR object tokenizer: sentence tokenizer object for the target language. Must have a method *split* that behaves like the one of MosesTokenizer. """ self.asr = asr self.tokenizer = tokenizer self.init() def init(self): """run this when starting or restarting processing""" self.audio_buffer = np.array([],dtype=np.float32) self.buffer_time_offset = 0 self.transcript_buffer = HypothesisBuffer() self.commited = [] self.last_chunked_at = 0 self.silence_iters = 0 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.last_chunked_at: 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 (commited) partial transcript. """ prompt, non_prompt = self.prompt() print("PROMPT:", prompt, file=sys.stderr) print("CONTEXT:", non_prompt, file=sys.stderr) print(f"transcribing {len(self.audio_buffer)/self.SAMPLING_RATE:2.2f} seconds from {self.buffer_time_offset:2.2f}",file=sys.stderr) 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) print(">>>>COMPLETE NOW:",self.to_flush(o),file=sys.stderr,flush=True) print("INCOMPLETE:",self.to_flush(self.transcript_buffer.complete()),file=sys.stderr,flush=True) # there is a newly confirmed text if o: # we trim all the completed sentences from the audio buffer self.chunk_completed_sentence() # ...segments could be considered #self.chunk_completed_segment(res) # # self.silence_iters = 0 # this was an attempt to trim silence/non-linguistic noise detected by the fact that Whisper doesn't transcribe anything for 3-times in a row. # It seemed not working better, or needs to be debugged. # elif self.transcript_buffer.complete(): # self.silence_iters = 0 # elif not self.transcript_buffer.complete(): # # print("NOT COMPLETE:",to_flush(self.transcript_buffer.complete()),file=sys.stderr,flush=True) # self.silence_iters += 1 # if self.silence_iters >= 3: # n = self.last_chunked_at ## self.chunk_completed_sentence() ## if n == self.last_chunked_at: # self.chunk_at(self.last_chunked_at+self.chunk) # print(f"\tCHUNK: 3-times silence! chunk_at {n}+{self.chunk}",file=sys.stderr) ## self.silence_iters = 0 # if the audio buffer is longer than 30s, trim it... if len(self.audio_buffer)/self.SAMPLING_RATE > 30: # ...on the last completed segment (labeled by Whisper) 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] print(f"chunking because of len",file=sys.stderr) #self.chunk_at(t) print(f"len of buffer now: {len(self.audio_buffer)/self.SAMPLING_RATE:2.2f}",file=sys.stderr) return self.to_flush(o) def chunk_completed_sentence(self): if self.commited == []: return print(self.commited,file=sys.stderr) sents = self.words_to_sentences(self.commited) for s in sents: print("\t\tSENT:",s,file=sys.stderr) 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] print(f"--- sentence chunked at {chunk_at:2.2f}",file=sys.stderr) 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: print(f"--- segment chunked at {e:2.2f}",file=sys.stderr) self.chunk_at(e) else: print(f"--- last segment not within commited area",file=sys.stderr) else: print(f"--- not enough segments to chunk",file=sys.stderr) 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 self.last_chunked_at = 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) 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) print("last, noncommited:",f,file=sys.stderr) 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) def create_tokenizer(lan): if lan == "uk": import tokenize_uk class UkrainianTokenizer: def split(self, text): return tokenize_uk.tokenize_sents(text) return UkrainianTokenizer() from mosestokenizer import MosesTokenizer return MosesTokenizer(lan) ## main: 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.") 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".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='en', help="Language code for transcription, e.g. en,de,cs.") parser.add_argument('--task', type=str, default='transcribe', choices=["transcribe","translate"],help="Transcribe or translate.") parser.add_argument('--start_at', type=float, default=0.0, help='Start processing audio at this time.') parser.add_argument('--backend', type=str, default="faster-whisper", choices=["faster-whisper", "whisper_timestamped"],help='Load only this backend for Whisper processing.') 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.') parser.add_argument('--vad', action="store_true", default=False, help='Use VAD = voice activity detection, with the default parameters.') args = parser.parse_args() if args.offline and args.comp_unaware: print("No or one option from --offline and --comp_unaware are available, not both. Exiting.",file=sys.stderr) sys.exit(1) audio_path = args.audio_path SAMPLING_RATE = 16000 duration = len(load_audio(audio_path))/SAMPLING_RATE print("Audio duration is: %2.2f seconds" % duration, file=sys.stderr) size = args.model language = args.lan t = time.time() print(f"Loading Whisper {size} model for {language}...",file=sys.stderr,end=" ",flush=True) #asr = WhisperASR(lan=language, modelsize=size) if args.backend == "faster-whisper": from faster_whisper import WhisperModel asr_cls = FasterWhisperASR else: import whisper import whisper_timestamped # from whisper_timestamped_model import WhisperTimestampedASR asr_cls = WhisperTimestampedASR asr = asr_cls(modelsize=size, lan=language, cache_dir=args.model_cache_dir, model_dir=args.model_dir) if args.task == "translate": asr.set_translate_task() tgt_language = "en" # Whisper translates into English else: tgt_language = language # Whisper transcribes in this language e = time.time() print(f"done. It took {round(e-t,2)} seconds.",file=sys.stderr) if args.vad: print("setting VAD filter",file=sys.stderr) asr.use_vad() min_chunk = args.min_chunk_size online = OnlineASRProcessor(asr,create_tokenizer(tgt_language)) # 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=sys.stderr,flush=True) print("%1.4f %1.0f %1.0f %s" % (now*1000, o[0]*1000,o[1]*1000,o[2]),flush=True) else: print(o,file=sys.stderr,flush=True) 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: print("assertion error",file=sys.stderr) pass 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: print("assertion error",file=sys.stderr) pass else: output_transcript(o, now=end) print(f"## last processed {end:.2f}s",file=sys.stderr,flush=True) beg = end end += min_chunk if end >= duration: break 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: print("assertion error",file=sys.stderr) pass else: output_transcript(o) now = time.time() - start print(f"## last processed {end:.2f} s, now is {now:.2f}, the latency is {now-end:.2f}",file=sys.stderr,flush=True) if end >= duration: break now = None o = online.finish() output_transcript(o, now=now)