From 2b8b245fa8f301e590a59bf24144e68fe3ca7678 Mon Sep 17 00:00:00 2001 From: makaveli10 Date: Tue, 22 Jul 2025 08:54:10 +0000 Subject: [PATCH 1/4] Add translation backend Translate from any language to any language with alirezamsh/small100 running in a thread and reading from a queue shared with transcription thread. Signed-off-by: makaveli10 --- README.md | 6 +- run_client.py | 12 +- whisper_live/backend/base.py | 24 +- .../backend/faster_whisper_backend.py | 6 +- whisper_live/backend/tokenization_small100.py | 365 ++++++++++++++++++ whisper_live/backend/translation_backend.py | 218 +++++++++++ whisper_live/client.py | 92 ++++- whisper_live/server.py | 45 ++- whisper_live/utils.py | 5 +- 9 files changed, 743 insertions(+), 30 deletions(-) create mode 100644 whisper_live/backend/tokenization_small100.py create mode 100644 whisper_live/backend/translation_backend.py diff --git a/README.md b/README.md index 7861b82..a571437 100644 --- a/README.md +++ b/README.md @@ -111,6 +111,8 @@ If you don't want this, set `--no_single_model`. - `save_output_recording`: Set to True to save the microphone input as a `.wav` file during live transcription. This option is helpful for recording sessions for later playback or analysis. Defaults to `False`. - `output_recording_filename`: Specifies the `.wav` file path where the microphone input will be saved if `save_output_recording` is set to `True`. - `mute_audio_playback`: Whether to mute audio playback when transcribing an audio file. Defaults to False. + - `enable_translation`: Start translation thread on the server (from any to any). + - `target_language`: Server translation thread's target translation language. ```python from whisper_live.client import TranscriptionClient @@ -124,6 +126,8 @@ client = TranscriptionClient( save_output_recording=True, # Only used for microphone input, False by Default output_recording_filename="./output_recording.wav", # Only used for microphone input mute_audio_playback=False, # Only used for file input, False by Default + enable_translation=True, + target_language="hi", ) ``` It connects to the server running on localhost at port 9090. Using a multilingual model, language for the transcription will be automatically detected. You can also use the language option to specify the target language for the transcription, in this case, English ("en"). The translate option should be set to `True` if we want to translate from the source language to English and `False` if we want to transcribe in the source language. @@ -195,7 +199,7 @@ Refer to [`ios-client`](https://github.com/collabora/WhisperLive/tree/main/Audio ``` ## Future Work -- [ ] Add translation to other languages on top of transcription. +- [x] Add translation to other languages on top of transcription. ## Blog Posts - [Transforming speech technology with WhisperLive](https://www.collabora.com/news-and-blog/blog/2024/05/28/transforming-speech-technology-with-whisperlive/) diff --git a/run_client.py b/run_client.py index 33035fb..827ebca 100644 --- a/run_client.py +++ b/run_client.py @@ -39,7 +39,15 @@ if __name__ == '__main__': help='Mute audio playback during transcription.') parser.add_argument('--save_output_recording', '-r', action='store_true', - help='Save the output recording, only used for microphone input.') + help='Save the output recording, only used for microphone input.') + parser.add_argument('--enable_translation', + action='store_true', + help='Enable translation of the transcription output.') + parser.add_argument('--target_language', '-tl', + type=str, + default='fr', + help='Target language for translation, e.g., "fr" for French.') + args = parser.parse_args() # Validate audio files @@ -70,5 +78,7 @@ if __name__ == '__main__': save_output_recording=args.save_output_recording, # Only used for microphone input, False by Default output_recording_filename=args.output_file, # Only used for microphone input mute_audio_playback=args.mute_audio_playback, # Only used for file input, False by Default + enable_translation=args.enable_translation, # Enable translation of the transcription output + target_language=args.target_language, # Target language for translation, e.g., "fr ) client(f) diff --git a/whisper_live/backend/base.py b/whisper_live/backend/base.py index 8d5a611..e45e503 100644 --- a/whisper_live/backend/base.py +++ b/whisper_live/backend/base.py @@ -2,6 +2,7 @@ import json import logging import threading import time +import queue import numpy as np @@ -31,6 +32,7 @@ class ServeClientBase(object): no_speech_thresh=0.45, clip_audio=False, same_output_threshold=10, + translation_queue=None, ): self.client_uid = client_uid self.websocket = websocket @@ -50,6 +52,7 @@ class ServeClientBase(object): self.same_output_count = 0 self.transcript = [] self.end_time_for_same_output = None + self.translation_queue = translation_queue # threading self.lock = threading.Lock() @@ -307,7 +310,14 @@ class ServeClientBase(object): continue if self.get_segment_no_speech_prob(s) > self.no_speech_thresh: continue - self.transcript.append(self.format_segment(start, end, text_, completed=True)) + completed_segment = self.format_segment(start, end, text_, completed=True) + self.transcript.append(completed_segment) + + if self.translation_queue: + try: + self.translation_queue.put(completed_segment.copy(), timeout=0.1) + except queue.Full: + logging.warning("Translation queue is full, skipping segment") offset = min(duration, self.get_segment_end(s)) # Process the last segment if its no_speech_prob is acceptable. @@ -340,12 +350,20 @@ class ServeClientBase(object): if not self.text or self.text[-1].strip().lower() != self.current_out.strip().lower(): self.text.append(self.current_out) with self.lock: - self.transcript.append(self.format_segment( + completed_segment = self.format_segment( self.timestamp_offset, self.timestamp_offset + min(duration, self.end_time_for_same_output), self.current_out, completed=True - )) + ) + self.transcript.append(completed_segment) + + if self.translation_queue: + try: + self.translation_queue.put(completed_segment.copy(), timeout=0.1) + except queue.Full: + logging.warning("Translation queue is full, skipping segment") + self.current_out = '' offset = min(duration, self.end_time_for_same_output) self.same_output_count = 0 diff --git a/whisper_live/backend/faster_whisper_backend.py b/whisper_live/backend/faster_whisper_backend.py index e2384c8..8df189e 100644 --- a/whisper_live/backend/faster_whisper_backend.py +++ b/whisper_live/backend/faster_whisper_backend.py @@ -30,8 +30,9 @@ class ServeClientFasterWhisper(ServeClientBase): send_last_n_segments=10, no_speech_thresh=0.45, clip_audio=False, - same_output_threshold=10, - cache_path="~/.cache/whisper-live/" + same_output_threshold=7, + cache_path="~/.cache/whisper-live/", + translation_queue=None, ): """ Initialize a ServeClient instance. @@ -61,6 +62,7 @@ class ServeClientFasterWhisper(ServeClientBase): no_speech_thresh, clip_audio, same_output_threshold, + translation_queue ) self.cache_path = cache_path self.model_sizes = [ diff --git a/whisper_live/backend/tokenization_small100.py b/whisper_live/backend/tokenization_small100.py new file mode 100644 index 0000000..f618d7b --- /dev/null +++ b/whisper_live/backend/tokenization_small100.py @@ -0,0 +1,365 @@ +# Copyright (c) 2022 Idiap Research Institute, http://www.idiap.ch/ +# Written by Alireza Mohammadshahi +# This is a modified version of https://github.com/huggingface/transformers/blob/main/src/transformers/models/m2m_100/tokenization_m2m_100.py +# which owns by Fariseq Authors and The HuggingFace Inc. team. +# +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Tokenization classes for SMALL100.""" +import json +import os +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, List, Optional, Tuple, Union + +import sentencepiece + +from transformers.tokenization_utils import BatchEncoding, PreTrainedTokenizer +from transformers.utils import logging + + +logger = logging.get_logger(__name__) + +SPIECE_UNDERLINE = "▁" + +VOCAB_FILES_NAMES = { + "vocab_file": "vocab.json", + "spm_file": "sentencepiece.bpe.model", + "tokenizer_config_file": "tokenizer_config.json", +} + +PRETRAINED_VOCAB_FILES_MAP = { + "vocab_file": { + "alirezamsh/small100": "https://huggingface.co/alirezamsh/small100/resolve/main/vocab.json", + }, + "spm_file": { + "alirezamsh/small100": "https://huggingface.co/alirezamsh/small100/resolve/main/sentencepiece.bpe.model", + }, + "tokenizer_config_file": { + "alirezamsh/small100": "https://huggingface.co/alirezamsh/small100/resolve/main/tokenizer_config.json", + }, +} + +PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = { + "alirezamsh/small100": 1024, +} + +# fmt: off +FAIRSEQ_LANGUAGE_CODES = { + "m2m100": ["af", "am", "ar", "ast", "az", "ba", "be", "bg", "bn", "br", "bs", "ca", "ceb", "cs", "cy", "da", "de", "el", "en", "es", "et", "fa", "ff", "fi", "fr", "fy", "ga", "gd", "gl", "gu", "ha", "he", "hi", "hr", "ht", "hu", "hy", "id", "ig", "ilo", "is", "it", "ja", "jv", "ka", "kk", "km", "kn", "ko", "lb", "lg", "ln", "lo", "lt", "lv", "mg", "mk", "ml", "mn", "mr", "ms", "my", "ne", "nl", "no", "ns", "oc", "or", "pa", "pl", "ps", "pt", "ro", "ru", "sd", "si", "sk", "sl", "so", "sq", "sr", "ss", "su", "sv", "sw", "ta", "th", "tl", "tn", "tr", "uk", "ur", "uz", "vi", "wo", "xh", "yi", "yo", "zh", "zu"] +} +# fmt: on + + +class SMALL100Tokenizer(PreTrainedTokenizer): + """ + Construct an SMALL100 tokenizer. Based on [SentencePiece](https://github.com/google/sentencepiece). + This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to + this superclass for more information regarding those methods. + Args: + vocab_file (`str`): + Path to the vocabulary file. + spm_file (`str`): + Path to [SentencePiece](https://github.com/google/sentencepiece) file (generally has a .spm extension) that + contains the vocabulary. + tgt_lang (`str`, *optional*): + A string representing the target language. + eos_token (`str`, *optional*, defaults to `""`): + The end of sequence token. + sep_token (`str`, *optional*, defaults to `""`): + The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for + sequence classification or for a text and a question for question answering. It is also used as the last + token of a sequence built with special tokens. + unk_token (`str`, *optional*, defaults to `""`): + The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this + token instead. + pad_token (`str`, *optional*, defaults to `""`): + The token used for padding, for example when batching sequences of different lengths. + language_codes (`str`, *optional*): + What language codes to use. Should be `"m2m100"`. + sp_model_kwargs (`dict`, *optional*): + Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for + SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things, + to set: + - `enable_sampling`: Enable subword regularization. + - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout. + - `nbest_size = {0,1}`: No sampling is performed. + - `nbest_size > 1`: samples from the nbest_size results. + - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice) + using forward-filtering-and-backward-sampling algorithm. + - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for + BPE-dropout. + Examples: + ```python + >>> from tokenization_small100 import SMALL100Tokenizer + >>> tokenizer = SMALL100Tokenizer.from_pretrained("alirezamsh/small100", tgt_lang="ro") + >>> src_text = " UN Chief Says There Is No Military Solution in Syria" + >>> tgt_text = "Şeful ONU declară că nu există o soluţie militară în Siria" + >>> model_inputs = tokenizer(src_text, text_target=tgt_text, return_tensors="pt") + >>> model(**model_inputs) # should work + ```""" + + vocab_files_names = VOCAB_FILES_NAMES + max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES + pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP + model_input_names = ["input_ids", "attention_mask"] + + prefix_tokens: List[int] = [] + suffix_tokens: List[int] = [] + + def __init__( + self, + vocab_file, + spm_file, + tgt_lang=None, + bos_token="", + eos_token="", + sep_token="", + pad_token="", + unk_token="", + language_codes="m2m100", + sp_model_kwargs: Optional[Dict[str, Any]] = None, + num_madeup_words=8, + **kwargs, + ) -> None: + self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs + + self.language_codes = language_codes + fairseq_language_code = FAIRSEQ_LANGUAGE_CODES[language_codes] + self.lang_code_to_token = {lang_code: f"__{lang_code}__" for lang_code in fairseq_language_code} + + kwargs["additional_special_tokens"] = kwargs.get("additional_special_tokens", []) + kwargs["additional_special_tokens"] += [ + self.get_lang_token(lang_code) + for lang_code in fairseq_language_code + if self.get_lang_token(lang_code) not in kwargs["additional_special_tokens"] + ] + + self.vocab_file = vocab_file + self.encoder = load_json(vocab_file) + self.decoder = {v: k for k, v in self.encoder.items()} + self.spm_file = spm_file + self.sp_model = load_spm(spm_file, self.sp_model_kwargs) + + self.encoder_size = len(self.encoder) + + self.lang_token_to_id = { + self.get_lang_token(lang_code): self.encoder_size + i for i, lang_code in enumerate(fairseq_language_code) + } + self.lang_code_to_id = {lang_code: self.encoder_size + i for i, lang_code in enumerate(fairseq_language_code)} + self.id_to_lang_token = {v: k for k, v in self.lang_token_to_id.items()} + + self._tgt_lang = tgt_lang if tgt_lang is not None else "en" + self.cur_lang_id = self.get_lang_id(self._tgt_lang) + self.num_madeup_words = num_madeup_words + + super().__init__( + tgt_lang=tgt_lang, + bos_token=bos_token, + eos_token=eos_token, + sep_token=sep_token, + unk_token=unk_token, + pad_token=pad_token, + language_codes=language_codes, + sp_model_kwargs=self.sp_model_kwargs, + num_madeup_words=num_madeup_words, + **kwargs, + ) + + self.set_lang_special_tokens(self._tgt_lang) + + + @property + def vocab_size(self) -> int: + return len(self.encoder) + len(self.lang_token_to_id) + self.num_madeup_words + + @property + def tgt_lang(self) -> str: + return self._tgt_lang + + @tgt_lang.setter + def tgt_lang(self, new_tgt_lang: str) -> None: + self._tgt_lang = new_tgt_lang + self.set_lang_special_tokens(self._tgt_lang) + + def _tokenize(self, text: str) -> List[str]: + return self.sp_model.encode(text, out_type=str) + + def _convert_token_to_id(self, token): + if token in self.lang_token_to_id: + return self.lang_token_to_id[token] + return self.encoder.get(token, self.encoder[self.unk_token]) + + def _convert_id_to_token(self, index: int) -> str: + """Converts an index (integer) in a token (str) using the decoder.""" + if index in self.id_to_lang_token: + return self.id_to_lang_token[index] + return self.decoder.get(index, self.unk_token) + + def convert_tokens_to_string(self, tokens: List[str]) -> str: + """Converts a sequence of tokens (strings for sub-words) in a single string.""" + return self.sp_model.decode(tokens) + + def get_special_tokens_mask( + self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False + ) -> List[int]: + """ + Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding + special tokens using the tokenizer `prepare_for_model` method. + Args: + token_ids_0 (`List[int]`): + List of IDs. + token_ids_1 (`List[int]`, *optional*): + Optional second list of IDs for sequence pairs. + already_has_special_tokens (`bool`, *optional*, defaults to `False`): + Whether or not the token list is already formatted with special tokens for the model. + Returns: + `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token. + """ + + if already_has_special_tokens: + return super().get_special_tokens_mask( + token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True + ) + + prefix_ones = [1] * len(self.prefix_tokens) + suffix_ones = [1] * len(self.suffix_tokens) + if token_ids_1 is None: + return prefix_ones + ([0] * len(token_ids_0)) + suffix_ones + return prefix_ones + ([0] * len(token_ids_0)) + ([0] * len(token_ids_1)) + suffix_ones + + def build_inputs_with_special_tokens( + self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None + ) -> List[int]: + """ + Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and + adding special tokens. An MBART sequence has the following format, where `X` represents the sequence: + - `input_ids` (for encoder) `X [eos, src_lang_code]` + - `decoder_input_ids`: (for decoder) `X [eos, tgt_lang_code]` + BOS is never used. Pairs of sequences are not the expected use case, but they will be handled without a + separator. + Args: + token_ids_0 (`List[int]`): + List of IDs to which the special tokens will be added. + token_ids_1 (`List[int]`, *optional*): + Optional second list of IDs for sequence pairs. + Returns: + `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens. + """ + if token_ids_1 is None: + if self.prefix_tokens is None: + return token_ids_0 + self.suffix_tokens + else: + return self.prefix_tokens + token_ids_0 + self.suffix_tokens + # We don't expect to process pairs, but leave the pair logic for API consistency + if self.prefix_tokens is None: + return token_ids_0 + token_ids_1 + self.suffix_tokens + else: + return self.prefix_tokens + token_ids_0 + token_ids_1 + self.suffix_tokens + + def get_vocab(self) -> Dict: + vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)} + vocab.update(self.added_tokens_encoder) + return vocab + + def __getstate__(self) -> Dict: + state = self.__dict__.copy() + state["sp_model"] = None + return state + + def __setstate__(self, d: Dict) -> None: + self.__dict__ = d + + # for backward compatibility + if not hasattr(self, "sp_model_kwargs"): + self.sp_model_kwargs = {} + + self.sp_model = load_spm(self.spm_file, self.sp_model_kwargs) + + def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]: + save_dir = Path(save_directory) + if not save_dir.is_dir(): + raise OSError(f"{save_directory} should be a directory") + vocab_save_path = save_dir / ( + (filename_prefix + "-" if filename_prefix else "") + self.vocab_files_names["vocab_file"] + ) + spm_save_path = save_dir / ( + (filename_prefix + "-" if filename_prefix else "") + self.vocab_files_names["spm_file"] + ) + + save_json(self.encoder, vocab_save_path) + + if os.path.abspath(self.spm_file) != os.path.abspath(spm_save_path) and os.path.isfile(self.spm_file): + copyfile(self.spm_file, spm_save_path) + elif not os.path.isfile(self.spm_file): + with open(spm_save_path, "wb") as fi: + content_spiece_model = self.sp_model.serialized_model_proto() + fi.write(content_spiece_model) + + return (str(vocab_save_path), str(spm_save_path)) + + def prepare_seq2seq_batch( + self, + src_texts: List[str], + tgt_texts: Optional[List[str]] = None, + tgt_lang: str = "ro", + **kwargs, + ) -> BatchEncoding: + self.tgt_lang = tgt_lang + self.set_lang_special_tokens(self.tgt_lang) + return super().prepare_seq2seq_batch(src_texts, tgt_texts, **kwargs) + + def _build_translation_inputs(self, raw_inputs, tgt_lang: Optional[str], **extra_kwargs): + """Used by translation pipeline, to prepare inputs for the generate function""" + if tgt_lang is None: + raise ValueError("Translation requires a `tgt_lang` for this model") + self.tgt_lang = tgt_lang + inputs = self(raw_inputs, add_special_tokens=True, **extra_kwargs) + return inputs + + def _switch_to_input_mode(self): + self.set_lang_special_tokens(self.tgt_lang) + + def _switch_to_target_mode(self): + self.prefix_tokens = None + self.suffix_tokens = [self.eos_token_id] + + def set_lang_special_tokens(self, src_lang: str) -> None: + """Reset the special tokens to the tgt lang setting. No prefix and suffix=[eos, tgt_lang_code].""" + lang_token = self.get_lang_token(src_lang) + self.cur_lang_id = self.lang_token_to_id[lang_token] + self.prefix_tokens = [self.cur_lang_id] + self.suffix_tokens = [self.eos_token_id] + + def get_lang_token(self, lang: str) -> str: + return self.lang_code_to_token[lang] + + def get_lang_id(self, lang: str) -> int: + lang_token = self.get_lang_token(lang) + return self.lang_token_to_id[lang_token] + + +def load_spm(path: str, sp_model_kwargs: Dict[str, Any]) -> sentencepiece.SentencePieceProcessor: + spm = sentencepiece.SentencePieceProcessor(**sp_model_kwargs) + spm.Load(str(path)) + return spm + + +def load_json(path: str) -> Union[Dict, List]: + with open(path, "r") as f: + return json.load(f) + + +def save_json(data, path: str) -> None: + with open(path, "w") as f: + json.dump(data, f, indent=2) \ No newline at end of file diff --git a/whisper_live/backend/translation_backend.py b/whisper_live/backend/translation_backend.py new file mode 100644 index 0000000..837a6eb --- /dev/null +++ b/whisper_live/backend/translation_backend.py @@ -0,0 +1,218 @@ +import json +import logging +import threading +import time +import queue +from typing import Dict, Any, Optional +import torch +import threading +from transformers import M2M100ForConditionalGeneration +from whisper_live.backend.tokenization_small100 import SMALL100Tokenizer + +from whisper_live.backend.base import ServeClientBase + + +class ServeClientTranslation(ServeClientBase): + """ + Handles translation of completed transcription segments in a separate thread. + Reads from a queue populated by the transcription backend and sends translated + segments back to the client via WebSocket. + """ + + def __init__( + self, + client_uid, + websocket, + translation_queue, + target_language="fr", + send_last_n_segments=10, + model_name="alirezamsh/small100" + ): + """ + Initialize the translation client. + + Args: + client_uid (str): Unique identifier for the client + websocket: WebSocket connection to the client + translation_queue (queue.Queue): Queue containing completed segments to translate + target_language (str): Target language code (default: "fr" for French) + send_last_n_segments (int): Number of recent translated segments to send + model_name (str): Translation model name to use + """ + super().__init__(client_uid, websocket, send_last_n_segments) + self.translation_queue = translation_queue + self.target_language = target_language + self.model_name = model_name + self.translated_segments = [] + self.translation_model = None + self.tokenizer = None + self.device = None + self.model_loaded = False + self.load_translation_model() + + def load_translation_model(self): + """Load the translation model and tokenizer.""" + try: + self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") + logging.info(f"Loading translation model on device: {self.device}") + + self.translation_model = M2M100ForConditionalGeneration.from_pretrained( + self.model_name + ).to(self.device) + self.tokenizer = SMALL100Tokenizer.from_pretrained(self.model_name) + self.tokenizer.tgt_lang = self.target_language + + self.model_loaded = True + logging.info(f"Translation model loaded successfully. Target language: {self.target_language}") + except Exception as e: + logging.error(f"Failed to load translation model: {e}") + self.translation_model = None + self.tokenizer = None + self.model_loaded = False + + def translate_text(self, text: str) -> str: + """ + Translate a single text segment. + + Args: + text (str): Text to translate + + Returns: + str: Translated text or original text if translation fails + """ + if not self.model_loaded or not text.strip(): + return text + + try: + # Encode input and move to device + encoded_input = self.tokenizer(text, return_tensors="pt").to(self.device) + + # Generate translation + with torch.no_grad(): + generated_tokens = self.translation_model.generate(**encoded_input) + + # Decode output + output = self.tokenizer.batch_decode(generated_tokens, skip_special_tokens=True) + return output[0] if output else text + + except Exception as e: + logging.error(f"Translation failed for text '{text}': {e}") + return text + + def process_translation_queue(self): + """ + Process segments from the translation queue. + Continuously reads from the queue until None is received (exit signal). + """ + logging.info(f"Starting translation processing for client {self.client_uid}") + + while not self.exit: + try: + # Get segment from queue with timeout + segment = self.translation_queue.get(timeout=1.0) + + # Check for exit signal + if segment is None: + logging.info(f"Received exit signal for translation client {self.client_uid}") + break + + # Only translate completed segments + if not segment.get("completed", False): + self.translation_queue.task_done() + continue + + # Translate the segment + original_text = segment.get("text", "") + translated_text = self.translate_text(original_text) + + # Create translated segment + translated_segment = { + "start": segment["start"], + "end": segment["end"], + "text": translated_text, + "completed": segment.get("completed", False), + "target_language": self.target_language + } + + self.translated_segments.append(translated_segment) + segments_to_send = self.prepare_translated_segments() + self.send_translation_to_client(segments_to_send) + + self.translation_queue.task_done() + + except queue.Empty: + continue + except Exception as e: + logging.error(f"Error processing translation queue: {e}") + continue + + logging.info(f"Translation processing ended for client {self.client_uid}") + + def prepare_translated_segments(self): + """ + Prepare the last n translated segments to send to client. + + Returns: + list: List of recent translated segments + """ + if len(self.translated_segments) >= self.send_last_n_segments: + return self.translated_segments[-self.send_last_n_segments:] + return self.translated_segments[:] + + def send_translation_to_client(self, translated_segments): + """ + Send translated segments to the client via WebSocket. + + Args: + translated_segments (list): List of translated segments to send + """ + try: + self.websocket.send( + json.dumps({ + "uid": self.client_uid, + "translated_segments": translated_segments, + }) + ) + except Exception as e: + logging.error(f"[ERROR]: Sending translation data to client: {e}") + + def speech_to_text(self): + """ + Override parent method to handle translation processing. + This method will be called when the translation thread starts. + """ + self.process_translation_queue() + + def set_target_language(self, language: str): + """ + Change the target language for translation. + + Args: + language (str): New target language code + """ + self.target_language = language + if self.tokenizer: + self.tokenizer.tgt_lang = language + logging.info(f"Target language changed to: {language}") + + def cleanup(self): + """Clean up translation resources.""" + logging.info(f"Cleaning up translation resources for client {self.client_uid}") + self.exit = True + + try: + self.translation_queue.put(None, timeout=1.0) + except: + pass + + self.translated_segments.clear() + + if self.translation_model: + del self.translation_model + self.translation_model = None + if self.tokenizer: + del self.tokenizer + self.tokenizer = None + + if self.device and self.device.type == 'cuda': + torch.cuda.empty_cache() diff --git a/whisper_live/client.py b/whisper_live/client.py index ada8322..c740cd1 100644 --- a/whisper_live/client.py +++ b/whisper_live/client.py @@ -37,6 +37,10 @@ class Client: clip_audio=False, same_output_threshold=10, transcription_callback=None, + enable_translation=False, + target_language="fr", + translation_callback=None, + translation_srt_file_path="output_translated.srt", ): """ Initializes a Client instance for audio recording and streaming to a server. @@ -59,6 +63,10 @@ class Client: clip_audio (bool, optional): Whether to clip audio with no valid segments. Defaults to False. same_output_threshold (int, optional): Number of repeated outputs before considering it as a valid segment. Defaults to 10. transcription_callback (callable, optional): A callback function to handle transcription results. Default is None. + enable_translation (float, optional): Whether to enable translation from any to any language. Defaults to False. + target_language (str, optional): Target language for translation. Defaults to 'fr'. + translation_callback (callable, optional): A callback function to handle translation results. Default is None. + translation_srt_file_path (str, optional): The file path to save the translated output SRT file. Default is "output_translated.srt". """ self.recording = False self.task = "transcribe" @@ -81,6 +89,12 @@ class Client: self.same_output_threshold = same_output_threshold self.transcription_callback = transcription_callback + # Translation-specific attributes + self.enable_translation = enable_translation + self.target_language = target_language + self.translation_callback = translation_callback + self.translation_srt_file_path = translation_srt_file_path + self.last_translated_segment = None if translate: self.task = "translate" @@ -110,6 +124,7 @@ class Client: self.ws_thread.start() self.transcript = [] + self.translated_transcript = [] print("[INFO]: * recording") def handle_status_messages(self, message_data): @@ -124,36 +139,53 @@ class Client: elif status == "WARNING": print(f"Message from Server: {message_data['message']}") - def process_segments(self, segments): + def process_segments(self, segments, translated=False): """Processes transcript segments.""" text = [] for i, seg in enumerate(segments): if not text or text[-1] != seg["text"]: - text.append(seg["text"]) + text.append(seg["text"].strip()) if i == len(segments) - 1 and not seg.get("completed", False): self.last_segment = seg - elif (self.server_backend == "faster_whisper" and seg.get("completed", False) and - (not self.transcript or - float(seg['start']) >= float(self.transcript[-1]['end']))): - self.transcript.append(seg) + elif self.server_backend == "faster_whisper" and seg.get("completed", False): + if translated: + if (not self.translated_transcript or float(seg['start']) >= float(self.translated_transcript[-1]['end'])): + self.translated_transcript.append(seg) + else: + if (not self.transcript or float(seg['start']) >= float(self.transcript[-1]['end'])): + self.transcript.append(seg) # update last received segment and last valid response time - if self.last_received_segment is None or self.last_received_segment != segments[-1]["text"]: - self.last_response_received = time.time() - self.last_received_segment = segments[-1]["text"] + if not translated: + if self.last_received_segment is None or self.last_received_segment != segments[-1]["text"]: + self.last_response_received = time.time() + self.last_received_segment = segments[-1]["text"] # call the transcription callback if provided - if self.transcription_callback and callable(self.transcription_callback): - try: - self.transcription_callback(" ".join(text), segments) # string, list - except Exception as e: - print(f"[WARN] transcription_callback raised: {e}") - return + if translated: + if self.translation_callback and callable(self.translation_callback): + try: + self.translation_callback(" ".join(text), segments) # string, list + except Exception as e: + print(f"[WARN] translation_callback raised: {e}") + return + else: + if self.transcription_callback and callable(self.transcription_callback): + try: + self.transcription_callback(" ".join(text), segments) # string, list + except Exception as e: + print(f"[WARN] transcription_callback raised: {e}") + return if self.log_transcription: - # Truncate to last 3 entries for brevity. - text = text[-3:] + original_text = [seg["text"] for seg in self.transcript[-4:]] + if self.last_segment is not None and self.last_segment["text"] not in original_text: + original_text.append(self.last_segment["text"]) + utils.clear_screen() - utils.print_transcript(text) + utils.print_transcript(original_text) + if self.enable_translation: + print(f"\n\nTRANSLATION to {self.target_language}:") + utils.print_transcript([seg["text"] for seg in self.translated_transcript[-4:]], translated=True) def on_message(self, ws, message): """ @@ -199,6 +231,9 @@ class Client: if "segments" in message.keys(): self.process_segments(message["segments"]) + + if "translated_segments" in message.keys(): + self.process_segments(message["translated_segments"], translated=True) def on_error(self, ws, error): print(f"[ERROR] WebSocket Error: {error}") @@ -234,6 +269,8 @@ class Client: "no_speech_thresh": self.no_speech_thresh, "clip_audio": self.clip_audio, "same_output_threshold": self.same_output_threshold, + "enable_translation": self.enable_translation, + "target_language": self.target_language, } ) ) @@ -293,6 +330,9 @@ class Client: self.transcript.append(self.last_segment) utils.create_srt_file(self.transcript, output_path) + if self.enable_translation: + utils.create_srt_file(self.translated_transcript, self.translation_srt_file_path) + def wait_before_disconnect(self): """Waits a bit before disconnecting in order to process pending responses.""" assert self.last_response_received @@ -692,7 +732,7 @@ class TranscriptionClient(TranscriptionTeeClient): """ Client for handling audio transcription tasks via a single WebSocket connection. - Acts as a high-level client for audio transcription tasks using a WebSocket connection. It can be used + Acts as a high-level client for audio transcription tasksoutput_transcription_path using a WebSocket connection. It can be used to send audio data for transcription to a server and receive transcribed text segments. Args: @@ -712,6 +752,10 @@ class TranscriptionClient(TranscriptionTeeClient): clip_audio (bool, optional): Whether to clip audio with no valid segments. Defaults to False. same_output_threshold (int, optional): Number of repeated outputs before considering it as a valid segment. Defaults to 10. transcription_callback (callable, optional): A callback function to handle transcription results. Default is None. + enable_translation (float, optional): Whether to enable translation from any to any language. Defaults to False. + target_language (str, optional): Target language for translation. Defaults to 'fr'. + translation_callback (callable, optional): A callback function to handle translation results. Default is None. + translation_srt_file_path (str, optional): The file path to save the translated output SRT file. Default is "output_translated.srt". Attributes: client (Client): An instance of the underlying Client class responsible for handling the WebSocket connection. @@ -742,6 +786,10 @@ class TranscriptionClient(TranscriptionTeeClient): clip_audio=False, same_output_threshold=10, transcription_callback=None, + enable_translation=False, + target_language="fr", + translation_callback=None, + translation_srt_file_path="./output_translated.srt", ): self.client = Client( host, @@ -758,12 +806,18 @@ class TranscriptionClient(TranscriptionTeeClient): clip_audio=clip_audio, same_output_threshold=same_output_threshold, transcription_callback=transcription_callback, + enable_translation=enable_translation, + target_language=target_language, + translation_callback=translation_callback, + translation_srt_file_path=translation_srt_file_path, ) if save_output_recording and not output_recording_filename.endswith(".wav"): raise ValueError(f"Please provide a valid `output_recording_filename`: {output_recording_filename}") if not output_transcription_path.endswith(".srt"): raise ValueError(f"Please provide a valid `output_transcription_path`: {output_transcription_path}. The file extension should be `.srt`.") + if not translation_srt_file_path.endswith(".srt"): + raise ValueError(f"Please provide a valid `translation_srt_file_path`: {translation_srt_file_path}. The file extension should be `.srt`.") TranscriptionTeeClient.__init__( self, [self.client], diff --git a/whisper_live/server.py b/whisper_live/server.py index f58eaec..c11d1f8 100644 --- a/whisper_live/server.py +++ b/whisper_live/server.py @@ -1,6 +1,7 @@ import os import time import threading +import queue import json import functools import logging @@ -12,10 +13,10 @@ from websockets.sync.server import serve from websockets.exceptions import ConnectionClosed from whisper_live.vad import VoiceActivityDetector from whisper_live.backend.base import ServeClientBase +from whisper_live.backend.translation_backend import ServeClientTranslation logging.basicConfig(level=logging.INFO) - class ClientManager: def __init__(self, max_clients=4, max_connection_time=600): """ @@ -157,6 +158,34 @@ class TranscriptionServer: ): client: Optional[ServeClientBase] = None + # Check if client wants translation + enable_translation = options.get("enable_translation", False) + target_language = options.get("target_language", "fr") + + # Create translation queue if translation is enabled + translation_queue = None + translation_client = None + translation_thread = None + + if enable_translation: + translation_queue = queue.Queue() + translation_client = ServeClientTranslation( + client_uid=options["uid"], + websocket=websocket, + translation_queue=translation_queue, + target_language=target_language, + send_last_n_segments=options.get("send_last_n_segments", 10) + ) + + # Start translation thread + translation_thread = threading.Thread( + target=translation_client.speech_to_text, + daemon=True + ) + translation_thread.start() + + logging.info(f"Translation enabled for client {options['uid']} with target language: {target_language}") + if self.backend.is_tensorrt(): try: from whisper_live.backend.trt_backend import ServeClientTensorRT @@ -235,6 +264,7 @@ class TranscriptionServer: clip_audio=options.get("clip_audio", False), same_output_threshold=options.get("same_output_threshold", 10), cache_path=self.cache_path, + translation_queue=translation_queue ) logging.info("Running faster_whisper backend.") @@ -245,6 +275,10 @@ class TranscriptionServer: if client is None: raise ValueError(f"Backend type {self.backend.value} not recognised or not handled.") + if translation_client: + client.translation_client = translation_client + client.translation_thread = translation_thread + self.client_manager.add_client(websocket, client) def get_audio_from_websocket(self, websocket): @@ -443,6 +477,13 @@ class TranscriptionServer: Args: websocket: The websocket associated with the client to be cleaned up. """ - if self.client_manager.get_client(websocket): + client = self.client_manager.get_client(websocket) + if client: + if hasattr(client, 'translation_client') and client.translation_client: + client.translation_client.cleanup() + + # Wait for translation thread to finish + if hasattr(client, 'translation_thread') and client.translation_thread: + client.translation_thread.join(timeout=2.0) self.client_manager.remove_client(websocket) diff --git a/whisper_live/utils.py b/whisper_live/utils.py index 1f9b2ad..a846d32 100644 --- a/whisper_live/utils.py +++ b/whisper_live/utils.py @@ -11,10 +11,11 @@ def clear_screen(): os.system("cls" if os.name == "nt" else "clear") -def print_transcript(text): +def print_transcript(text, translated=False): """Prints formatted transcript text.""" wrapper = textwrap.TextWrapper(width=60) - for line in wrapper.wrap(text="".join(text)): + text=" ".join(text) if translated else "".join(text) + for line in wrapper.wrap(text=text): print(line) From 39dfd7521fe2d185da6bebfd725f5ec0ea2f061a Mon Sep 17 00:00:00 2001 From: makaveli10 Date: Tue, 22 Jul 2025 09:05:54 +0000 Subject: [PATCH 2/4] Update requirements Signed-off-by: makaveli10 --- requirements/server.txt | 1 + 1 file changed, 1 insertion(+) diff --git a/requirements/server.txt b/requirements/server.txt index 773fe7e..76319ad 100644 --- a/requirements/server.txt +++ b/requirements/server.txt @@ -12,6 +12,7 @@ numpy<2 openai-whisper==20240930 tokenizers==0.20.3 transformers[torch] +sentencepiece # openvino librosa From 5ce401d4c63987e05eeafdb8d0505fa40c70158b Mon Sep 17 00:00:00 2001 From: makaveli10 Date: Tue, 22 Jul 2025 09:06:31 +0000 Subject: [PATCH 3/4] Update test_client to expect translation args Signed-off-by: makaveli10 --- tests/test_client.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tests/test_client.py b/tests/test_client.py index b140fc5..2808648 100644 --- a/tests/test_client.py +++ b/tests/test_client.py @@ -53,6 +53,8 @@ class TestClientCallbacks(BaseTestCase): "no_speech_thresh": 0.45, "clip_audio": False, "same_output_threshold": 10, + "enable_translation": False, + "target_language": "fr", }) self.client.on_open(self.mock_ws_app) self.mock_ws_app.send.assert_called_with(expected_message) From 04db67170b1651c5596722e5872460ac227d4809 Mon Sep 17 00:00:00 2001 From: makaveli10 Date: Tue, 22 Jul 2025 15:47:06 +0000 Subject: [PATCH 4/4] ServeClientTranslation import only when enable_tranlsation is True Signed-off-by: makaveli10 --- whisper_live/server.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/whisper_live/server.py b/whisper_live/server.py index c11d1f8..47a4543 100644 --- a/whisper_live/server.py +++ b/whisper_live/server.py @@ -13,7 +13,6 @@ from websockets.sync.server import serve from websockets.exceptions import ConnectionClosed from whisper_live.vad import VoiceActivityDetector from whisper_live.backend.base import ServeClientBase -from whisper_live.backend.translation_backend import ServeClientTranslation logging.basicConfig(level=logging.INFO) @@ -160,7 +159,6 @@ class TranscriptionServer: # Check if client wants translation enable_translation = options.get("enable_translation", False) - target_language = options.get("target_language", "fr") # Create translation queue if translation is enabled translation_queue = None @@ -168,7 +166,9 @@ class TranscriptionServer: translation_thread = None if enable_translation: + target_language = options.get("target_language", "fr") translation_queue = queue.Queue() + from whisper_live.backend.translation_backend import ServeClientTranslation translation_client = ServeClientTranslation( client_uid=options["uid"], websocket=websocket,