First commit of files

whisper_live/__version__.py

@@ -0,0 +1 @@
+__version__ = "0.4.1"

whisper_live/tensorrt_utils.py

@@ -0,0 +1,365 @@
+# SPDX-FileCopyrightText: Copyright (c) 2022-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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.
+import logging
+import os
+from collections import defaultdict
+from functools import lru_cache
+from pathlib import Path
+from subprocess import CalledProcessError, run
+from typing import Dict, Iterable, List, Optional, TextIO, Tuple, Union
+
+import kaldialign
+import numpy as np
+import soundfile
+import torch
+import torch.nn.functional as F
+
+Pathlike = Union[str, Path]
+
+SAMPLE_RATE = 16000
+N_FFT = 400
+HOP_LENGTH = 160
+CHUNK_LENGTH = 30
+N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE  # 480000 samples in a 30-second chunk
+
+
+def load_audio(file: str, sr: int = SAMPLE_RATE):
+    """
+    Open an audio file and read as mono waveform, resampling as necessary
+
+    Parameters
+    ----------
+    file: str
+        The audio file to open
+
+    sr: int
+        The sample rate to resample the audio if necessary
+
+    Returns
+    -------
+    A NumPy array containing the audio waveform, in float32 dtype.
+    """
+
+    # This launches a subprocess to decode audio while down-mixing
+    # and resampling as necessary.  Requires the ffmpeg CLI in PATH.
+    # fmt: off
+    cmd = [
+        "ffmpeg", "-nostdin", "-threads", "0", "-i", file, "-f", "s16le", "-ac",
+        "1", "-acodec", "pcm_s16le", "-ar",
+        str(sr), "-"
+    ]
+    # fmt: on
+    try:
+        out = run(cmd, capture_output=True, check=True).stdout
+    except CalledProcessError as e:
+        raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e
+
+    return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0
+
+
+def load_audio_wav_format(wav_path):
+    # make sure audio in .wav format
+    assert wav_path.endswith(
+        '.wav'), f"Only support .wav format, but got {wav_path}"
+    waveform, sample_rate = soundfile.read(wav_path)
+    assert sample_rate == 16000, f"Only support 16k sample rate, but got {sample_rate}"
+    return waveform, sample_rate
+
+
+def pad_or_trim(array, length: int = N_SAMPLES, *, axis: int = -1):
+    """
+    Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
+    """
+    if torch.is_tensor(array):
+        if array.shape[axis] > length:
+            array = array.index_select(dim=axis,
+                                       index=torch.arange(length,
+                                                          device=array.device))
+
+        if array.shape[axis] < length:
+            pad_widths = [(0, 0)] * array.ndim
+            pad_widths[axis] = (0, length - array.shape[axis])
+            array = F.pad(array,
+                          [pad for sizes in pad_widths[::-1] for pad in sizes])
+    else:
+        if array.shape[axis] > length:
+            array = array.take(indices=range(length), axis=axis)
+
+        if array.shape[axis] < length:
+            pad_widths = [(0, 0)] * array.ndim
+            pad_widths[axis] = (0, length - array.shape[axis])
+            array = np.pad(array, pad_widths)
+
+    return array
+
+
+@lru_cache(maxsize=None)
+def mel_filters(device,
+                n_mels: int,
+                mel_filters_dir: str = None) -> torch.Tensor:
+    """
+    load the mel filterbank matrix for projecting STFT into a Mel spectrogram.
+    Allows decoupling librosa dependency; saved using:
+
+        np.savez_compressed(
+            "mel_filters.npz",
+            mel_80=librosa.filters.mel(sr=16000, n_fft=400, n_mels=80),
+        )
+    """
+    assert n_mels in {80, 128}, f"Unsupported n_mels: {n_mels}"
+    if mel_filters_dir is None:
+        mel_filters_path = os.path.join(os.path.dirname(__file__), "assets",
+                                        "mel_filters.npz")
+    else:
+        mel_filters_path = os.path.join(mel_filters_dir, "mel_filters.npz")
+    with np.load(mel_filters_path) as f:
+        return torch.from_numpy(f[f"mel_{n_mels}"]).to(device)
+
+
+def log_mel_spectrogram(
+    audio: Union[str, np.ndarray, torch.Tensor],
+    n_mels: int,
+    padding: int = 0,
+    device: Optional[Union[str, torch.device]] = None,
+    return_duration: bool = False,
+    mel_filters_dir: str = None,
+):
+    """
+    Compute the log-Mel spectrogram of
+
+    Parameters
+    ----------
+    audio: Union[str, np.ndarray, torch.Tensor], shape = (*)
+        The path to audio or either a NumPy array or Tensor containing the audio waveform in 16 kHz
+
+    n_mels: int
+        The number of Mel-frequency filters, only 80 and 128 are supported
+
+    padding: int
+        Number of zero samples to pad to the right
+
+    device: Optional[Union[str, torch.device]]
+        If given, the audio tensor is moved to this device before STFT
+
+    Returns
+    -------
+    torch.Tensor, shape = (80 or 128, n_frames)
+        A Tensor that contains the Mel spectrogram
+    """
+    if not torch.is_tensor(audio):
+        if isinstance(audio, str):
+            if audio.endswith('.wav'):
+                audio, _ = load_audio_wav_format(audio)
+            else:
+                audio = load_audio(audio)
+        assert isinstance(audio,
+                          np.ndarray), f"Unsupported audio type: {type(audio)}"
+        duration = audio.shape[-1] / SAMPLE_RATE
+        audio = pad_or_trim(audio, N_SAMPLES)
+        audio = audio.astype(np.float32)
+        audio = torch.from_numpy(audio)
+
+    if device is not None:
+        audio = audio.to(device)
+    if padding > 0:
+        audio = F.pad(audio, (0, padding))
+    window = torch.hann_window(N_FFT).to(audio.device)
+    stft = torch.stft(audio,
+                      N_FFT,
+                      HOP_LENGTH,
+                      window=window,
+                      return_complex=True)
+    magnitudes = stft[..., :-1].abs()**2
+
+    filters = mel_filters(audio.device, n_mels, mel_filters_dir)
+    mel_spec = filters @ magnitudes
+
+    log_spec = torch.clamp(mel_spec, min=1e-10).log10()
+    log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
+    log_spec = (log_spec + 4.0) / 4.0
+    if return_duration:
+        return log_spec, duration
+    else:
+        return log_spec
+
+
+def store_transcripts(filename: Pathlike, texts: Iterable[Tuple[str, str,
+                                                                str]]) -> None:
+    """Save predicted results and reference transcripts to a file.
+    https://github.com/k2-fsa/icefall/blob/master/icefall/utils.py
+    Args:
+      filename:
+        File to save the results to.
+      texts:
+        An iterable of tuples. The first element is the cur_id, the second is
+        the reference transcript and the third element is the predicted result.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w") as f:
+        for cut_id, ref, hyp in texts:
+            print(f"{cut_id}:\tref={ref}", file=f)
+            print(f"{cut_id}:\thyp={hyp}", file=f)
+
+
+def write_error_stats(                                              # noqa: C901
+    f: TextIO,
+    test_set_name: str,
+    results: List[Tuple[str, str]],
+    enable_log: bool = True,
+) -> float:
+    """Write statistics based on predicted results and reference transcripts.
+    https://github.com/k2-fsa/icefall/blob/master/icefall/utils.py
+    It will write the following to the given file:
+
+        - WER
+        - number of insertions, deletions, substitutions, corrects and total
+          reference words. For example::
+
+              Errors: 23 insertions, 57 deletions, 212 substitutions, over 2606
+              reference words (2337 correct)
+
+        - The difference between the reference transcript and predicted result.
+          An instance is given below::
+
+            THE ASSOCIATION OF (EDISON->ADDISON) ILLUMINATING COMPANIES
+
+          The above example shows that the reference word is `EDISON`,
+          but it is predicted to `ADDISON` (a substitution error).
+
+          Another example is::
+
+            FOR THE FIRST DAY (SIR->*) I THINK
+
+          The reference word `SIR` is missing in the predicted
+          results (a deletion error).
+      results:
+        An iterable of tuples. The first element is the cur_id, the second is
+        the reference transcript and the third element is the predicted result.
+      enable_log:
+        If True, also print detailed WER to the console.
+        Otherwise, it is written only to the given file.
+    Returns:
+      Return None.
+    """
+    subs: Dict[Tuple[str, str], int] = defaultdict(int)
+    ins: Dict[str, int] = defaultdict(int)
+    dels: Dict[str, int] = defaultdict(int)
+
+    # `words` stores counts per word, as follows:
+    #   corr, ref_sub, hyp_sub, ins, dels
+    words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0])
+    num_corr = 0
+    ERR = "*"
+    for cut_id, ref, hyp in results:
+        ali = kaldialign.align(ref, hyp, ERR)
+        for ref_word, hyp_word in ali:
+            if ref_word == ERR:
+                ins[hyp_word] += 1
+                words[hyp_word][3] += 1
+            elif hyp_word == ERR:
+                dels[ref_word] += 1
+                words[ref_word][4] += 1
+            elif hyp_word != ref_word:
+                subs[(ref_word, hyp_word)] += 1
+                words[ref_word][1] += 1
+                words[hyp_word][2] += 1
+            else:
+                words[ref_word][0] += 1
+                num_corr += 1
+    ref_len = sum([len(r) for _, r, _ in results])
+    sub_errs = sum(subs.values())
+    ins_errs = sum(ins.values())
+    del_errs = sum(dels.values())
+    tot_errs = sub_errs + ins_errs + del_errs
+    tot_err_rate = "%.2f" % (100.0 * tot_errs / ref_len)
+
+    if enable_log:
+        logging.info(f"[{test_set_name}] %WER {tot_errs / ref_len:.2%} "
+                     f"[{tot_errs} / {ref_len}, {ins_errs} ins, "
+                     f"{del_errs} del, {sub_errs} sub ]")
+
+    print(f"%WER = {tot_err_rate}", file=f)
+    print(
+        f"Errors: {ins_errs} insertions, {del_errs} deletions, "
+        f"{sub_errs} substitutions, over {ref_len} reference "
+        f"words ({num_corr} correct)",
+        file=f,
+    )
+    print(
+        "Search below for sections starting with PER-UTT DETAILS:, "
+        "SUBSTITUTIONS:, DELETIONS:, INSERTIONS:, PER-WORD STATS:",
+        file=f,
+    )
+
+    print("", file=f)
+    print("PER-UTT DETAILS: corr or (ref->hyp)  ", file=f)
+    for cut_id, ref, hyp in results:
+        ali = kaldialign.align(ref, hyp, ERR)
+        combine_successive_errors = True
+        if combine_successive_errors:
+            ali = [[[x], [y]] for x, y in ali]
+            for i in range(len(ali) - 1):
+                if ali[i][0] != ali[i][1] and ali[i + 1][0] != ali[i + 1][1]:
+                    ali[i + 1][0] = ali[i][0] + ali[i + 1][0]
+                    ali[i + 1][1] = ali[i][1] + ali[i + 1][1]
+                    ali[i] = [[], []]
+            ali = [[
+                list(filter(lambda a: a != ERR, x)),
+                list(filter(lambda a: a != ERR, y)),
+            ] for x, y in ali]
+            ali = list(filter(lambda x: x != [[], []], ali))
+            ali = [[
+                ERR if x == [] else " ".join(x),
+                ERR if y == [] else " ".join(y),
+            ] for x, y in ali]
+
+        print(
+            f"{cut_id}:\t" + " ".join((ref_word if ref_word == hyp_word else
+                                       f"({ref_word}->{hyp_word})"
+                                       for ref_word, hyp_word in ali)),
+            file=f,
+        )
+
+    print("", file=f)
+    print("SUBSTITUTIONS: count ref -> hyp", file=f)
+
+    for count, (ref, hyp) in sorted([(v, k) for k, v in subs.items()],
+                                    reverse=True):
+        print(f"{count}   {ref} -> {hyp}", file=f)
+
+    print("", file=f)
+    print("DELETIONS: count ref", file=f)
+    for count, ref in sorted([(v, k) for k, v in dels.items()], reverse=True):
+        print(f"{count}   {ref}", file=f)
+
+    print("", file=f)
+    print("INSERTIONS: count hyp", file=f)
+    for count, hyp in sorted([(v, k) for k, v in ins.items()], reverse=True):
+        print(f"{count}   {hyp}", file=f)
+
+    print("", file=f)
+    print("PER-WORD STATS: word  corr tot_errs count_in_ref count_in_hyp",
+          file=f)
+    for _, word, counts in sorted([(sum(v[1:]), k, v)
+                                   for k, v in words.items()],
+                                  reverse=True):
+        (corr, ref_sub, hyp_sub, ins, dels) = counts
+        tot_errs = ref_sub + hyp_sub + ins + dels
+        ref_count = corr + ref_sub + dels
+        hyp_count = corr + hyp_sub + ins
+
+        print(f"{word}   {corr} {tot_errs} {ref_count} {hyp_count}", file=f)
+    return float(tot_err_rate)

whisper_live/transcriber_tensorrt.py

@@ -0,0 +1,320 @@
+import json
+import re
+from collections import OrderedDict
+from pathlib import Path
+from typing import Union
+
+import torch
+import numpy as np
+import torch.nn.functional as F
+from whisper.tokenizer import get_tokenizer
+from whisper_live.tensorrt_utils import (mel_filters, load_audio_wav_format, pad_or_trim, load_audio)
+
+import tensorrt_llm
+import tensorrt_llm.logger as logger
+from tensorrt_llm._utils import (str_dtype_to_torch, str_dtype_to_trt,
+                                 trt_dtype_to_torch)
+from tensorrt_llm.runtime import ModelConfig, SamplingConfig
+from tensorrt_llm.runtime.session import Session, TensorInfo
+
+
+SAMPLE_RATE = 16000
+N_FFT = 400
+HOP_LENGTH = 160
+CHUNK_LENGTH = 30
+N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE  # 480000 samples in a 30-second chunk
+
+
+class WhisperEncoding:
+
+    def __init__(self, engine_dir):
+        self.session = self.get_session(engine_dir)
+
+    def get_session(self, engine_dir):
+        config_path = engine_dir / 'encoder_config.json'
+        with open(config_path, 'r') as f:
+            config = json.load(f)
+
+        dtype = config['builder_config']['precision']
+        n_mels = config['builder_config']['n_mels']
+        num_languages = config['builder_config']['num_languages']
+
+        self.dtype = dtype
+        self.n_mels = n_mels
+        self.num_languages = num_languages
+
+        serialize_path = engine_dir / f'whisper_encoder_{self.dtype}_tp1_rank0.engine'
+
+        with open(serialize_path, 'rb') as f:
+            session = Session.from_serialized_engine(f.read())
+
+        return session
+
+    def get_audio_features(self, mel):
+        inputs = OrderedDict()
+        output_list = []
+
+        inputs.update({'x': mel})
+        output_list.append(
+            TensorInfo('x', str_dtype_to_trt(self.dtype), mel.shape))
+
+        output_info = (self.session).infer_shapes(output_list)
+
+        logger.debug(f'output info {output_info}')
+        outputs = {
+            t.name: torch.empty(tuple(t.shape),
+                                dtype=trt_dtype_to_torch(t.dtype),
+                                device='cuda')
+            for t in output_info
+        }
+        stream = torch.cuda.current_stream()
+        ok = self.session.run(inputs=inputs,
+                              outputs=outputs,
+                              stream=stream.cuda_stream)
+        assert ok, 'Engine execution failed'
+        stream.synchronize()
+        audio_features = outputs['output']
+        return audio_features
+
+
+class WhisperDecoding:
+
+    def __init__(self, engine_dir, runtime_mapping, debug_mode=False):
+
+        self.decoder_config = self.get_config(engine_dir)
+        self.decoder_generation_session = self.get_session(
+            engine_dir, runtime_mapping, debug_mode)
+
+    def get_config(self, engine_dir):
+        config_path = engine_dir / 'decoder_config.json'
+        with open(config_path, 'r') as f:
+            config = json.load(f)
+        decoder_config = OrderedDict()
+        decoder_config.update(config['plugin_config'])
+        decoder_config.update(config['builder_config'])
+        return decoder_config
+
+    def get_session(self, engine_dir, runtime_mapping, debug_mode=False):
+        dtype = self.decoder_config['precision']
+        serialize_path = engine_dir / f'whisper_decoder_{dtype}_tp1_rank0.engine'
+        with open(serialize_path, "rb") as f:
+            decoder_engine_buffer = f.read()
+
+        decoder_model_config = ModelConfig(
+            num_heads=self.decoder_config['num_heads'],
+            num_kv_heads=self.decoder_config['num_heads'],
+            hidden_size=self.decoder_config['hidden_size'],
+            vocab_size=self.decoder_config['vocab_size'],
+            num_layers=self.decoder_config['num_layers'],
+            gpt_attention_plugin=self.decoder_config['gpt_attention_plugin'],
+            remove_input_padding=self.decoder_config['remove_input_padding'],
+            cross_attention=self.decoder_config['cross_attention'],
+            has_position_embedding=self.
+            decoder_config['has_position_embedding'],
+            has_token_type_embedding=self.
+            decoder_config['has_token_type_embedding'],
+        )
+        decoder_generation_session = tensorrt_llm.runtime.GenerationSession(
+            decoder_model_config,
+            decoder_engine_buffer,
+            runtime_mapping,
+            debug_mode=debug_mode)
+
+        return decoder_generation_session
+
+    def generate(self,
+                 decoder_input_ids,
+                 encoder_outputs,
+                 eot_id,
+                 max_new_tokens=40,
+                 num_beams=1):
+        encoder_input_lengths = torch.tensor(
+            [encoder_outputs.shape[1] for x in range(encoder_outputs.shape[0])],
+            dtype=torch.int32,
+            device='cuda')
+
+        decoder_input_lengths = torch.tensor([
+            decoder_input_ids.shape[-1]
+            for _ in range(decoder_input_ids.shape[0])
+        ],
+                                             dtype=torch.int32,
+                                             device='cuda')
+        decoder_max_input_length = torch.max(decoder_input_lengths).item()
+
+        # generation config
+        sampling_config = SamplingConfig(end_id=eot_id,
+                                         pad_id=eot_id,
+                                         num_beams=num_beams)
+        self.decoder_generation_session.setup(
+            decoder_input_lengths.size(0),
+            decoder_max_input_length,
+            max_new_tokens,
+            beam_width=num_beams,
+            encoder_max_input_length=encoder_outputs.shape[1])
+
+        torch.cuda.synchronize()
+
+        decoder_input_ids = decoder_input_ids.type(torch.int32).cuda()
+        output_ids = self.decoder_generation_session.decode(
+            decoder_input_ids,
+            decoder_input_lengths,
+            sampling_config,
+            encoder_output=encoder_outputs,
+            encoder_input_lengths=encoder_input_lengths,
+        )
+        torch.cuda.synchronize()
+
+        # get the list of int from output_ids tensor
+        output_ids = output_ids.cpu().numpy().tolist()
+        return output_ids
+
+
+class WhisperTRTLLM(object):
+
+    def __init__(self, engine_dir, assets_dir=None, device=None, is_multilingual=False,
+                 language="en", task="transcribe"):
+        world_size = 1
+        runtime_rank = tensorrt_llm.mpi_rank()
+        runtime_mapping = tensorrt_llm.Mapping(world_size, runtime_rank)
+        torch.cuda.set_device(runtime_rank % runtime_mapping.gpus_per_node)
+        engine_dir = Path(engine_dir)
+
+        self.encoder = WhisperEncoding(engine_dir)
+        self.decoder = WhisperDecoding(engine_dir,
+                                       runtime_mapping,
+                                       debug_mode=False)
+        self.n_mels = self.encoder.n_mels
+        # self.tokenizer = get_tokenizer(num_languages=self.encoder.num_languages,
+        #                                tokenizer_dir=assets_dir)
+        self.device = device
+        self.tokenizer = get_tokenizer(
+            is_multilingual,
+            num_languages=self.encoder.num_languages,
+            language=language,
+            task=task,
+        )
+        self.filters = mel_filters(self.device, self.encoder.n_mels, assets_dir)
+
+    def log_mel_spectrogram(
+        self,
+        audio: Union[str, np.ndarray, torch.Tensor],
+        padding: int = 0,
+        return_duration=True
+    ):
+        """
+        Compute the log-Mel spectrogram of
+
+        Parameters
+        ----------
+        audio: Union[str, np.ndarray, torch.Tensor], shape = (*)
+            The path to audio or either a NumPy array or Tensor containing the audio waveform in 16 kHz
+
+        n_mels: int
+            The number of Mel-frequency filters, only 80 and 128 are supported
+
+        padding: int
+            Number of zero samples to pad to the right
+
+        device: Optional[Union[str, torch.device]]
+            If given, the audio tensor is moved to this device before STFT
+
+        Returns
+        -------
+        torch.Tensor, shape = (80 or 128, n_frames)
+            A Tensor that contains the Mel spectrogram
+        """
+        if not torch.is_tensor(audio):
+            if isinstance(audio, str):
+                if audio.endswith('.wav'):
+                    audio, _ = load_audio_wav_format(audio)
+                else:
+                    audio = load_audio(audio)
+            assert isinstance(audio, np.ndarray), f"Unsupported audio type: {type(audio)}"
+            duration = audio.shape[-1] / SAMPLE_RATE
+            audio = pad_or_trim(audio, N_SAMPLES)
+            audio = audio.astype(np.float32)
+            audio = torch.from_numpy(audio)
+
+        if self.device is not None:
+            audio = audio.to(self.device)
+        if padding > 0:
+            audio = F.pad(audio, (0, padding))
+        window = torch.hann_window(N_FFT).to(audio.device)
+        stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=True)
+        magnitudes = stft[..., :-1].abs()**2
+
+        mel_spec = self.filters @ magnitudes
+
+        log_spec = torch.clamp(mel_spec, min=1e-10).log10()
+        log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
+        log_spec = (log_spec + 4.0) / 4.0
+        if return_duration:
+            return log_spec, duration
+        else:
+            return log_spec
+
+    def process_batch(
+            self,
+            mel,
+            text_prefix="<|startoftranscript|><|en|><|transcribe|><|notimestamps|>",
+            num_beams=1):
+        prompt_id = self.tokenizer.encode(
+            text_prefix, allowed_special=set(self.tokenizer.special_tokens.keys()))
+
+        prompt_id = torch.tensor(prompt_id)
+        batch_size = mel.shape[0]
+        decoder_input_ids = prompt_id.repeat(batch_size, 1)
+
+        encoder_output = self.encoder.get_audio_features(mel)
+        output_ids = self.decoder.generate(decoder_input_ids,
+                                           encoder_output,
+                                           self.tokenizer.eot,
+                                           max_new_tokens=96,
+                                           num_beams=num_beams)
+        texts = []
+        for i in range(len(output_ids)):
+            text = self.tokenizer.decode(output_ids[i][0]).strip()
+            texts.append(text)
+        return texts
+
+    def transcribe(
+            self,
+            mel,
+            text_prefix="<|startoftranscript|><|en|><|transcribe|><|notimestamps|>",
+            dtype='float16',
+            batch_size=1,
+            num_beams=1,
+            ):
+        mel = mel.type(str_dtype_to_torch(dtype))
+        mel = mel.unsqueeze(0)
+        predictions = self.process_batch(mel, text_prefix, num_beams)
+        prediction = predictions[0]
+
+        # remove all special tokens in the prediction
+        prediction = re.sub(r'<\|.*?\|>', '', prediction)
+        return prediction.strip()
+
+
+def decode_wav_file(
+        model,
+        mel,
+        text_prefix="<|startoftranscript|><|en|><|transcribe|><|notimestamps|>",
+        dtype='float16',
+        batch_size=1,
+        num_beams=1,
+        normalizer=None,
+        mel_filters_dir=None):
+
+    mel = mel.type(str_dtype_to_torch(dtype))
+    mel = mel.unsqueeze(0)
+    # repeat the mel spectrogram to match the batch size
+    mel = mel.repeat(batch_size, 1, 1)
+    predictions = model.process_batch(mel, text_prefix, num_beams)
+    prediction = predictions[0]
+
+    # remove all special tokens in the prediction
+    prediction = re.sub(r'<\|.*?\|>', '', prediction)
+    if normalizer:
+        prediction = normalizer(prediction)
+
+    return prediction.strip()

whisper_live/utils.py

@@ -0,0 +1,82 @@
+import os
+import textwrap
+import scipy
+import numpy as np
+import av
+from pathlib import Path
+
+
+def clear_screen():
+    """Clears the console screen."""
+    os.system("cls" if os.name == "nt" else "clear")
+
+
+def print_transcript(text):
+    """Prints formatted transcript text."""
+    wrapper = textwrap.TextWrapper(width=60)
+    for line in wrapper.wrap(text="".join(text)):
+        print(line)
+
+
+def format_time(s):
+    """Convert seconds (float) to SRT time format."""
+    hours = int(s // 3600)
+    minutes = int((s % 3600) // 60)
+    seconds = int(s % 60)
+    milliseconds = int((s - int(s)) * 1000)
+    return f"{hours:02}:{minutes:02}:{seconds:02},{milliseconds:03}"
+
+
+def create_srt_file(segments, resampled_file):
+    with open(resampled_file, 'w', encoding='utf-8') as srt_file:
+        segment_number = 1
+        for segment in segments:
+            start_time = format_time(float(segment['start']))
+            end_time = format_time(float(segment['end']))
+            text = segment['text']
+
+            srt_file.write(f"{segment_number}\n")
+            srt_file.write(f"{start_time} --> {end_time}\n")
+            srt_file.write(f"{text}\n\n")
+
+            segment_number += 1
+
+
+def resample(file: str, sr: int = 16000):
+    """
+    Resample the audio file to 16kHz.
+
+    Args:
+        file (str): The audio file to open
+        sr (int): The sample rate to resample the audio if necessary
+
+    Returns:
+        resampled_file (str): The resampled audio file
+    """
+    container = av.open(file)
+    stream = next(s for s in container.streams if s.type == 'audio')
+
+    resampler = av.AudioResampler(
+        format='s16',
+        layout='mono',
+        rate=sr,
+    )
+
+    resampled_file = Path(file).stem + "_resampled.wav"
+    output_container = av.open(resampled_file, mode='w')
+    output_stream = output_container.add_stream('pcm_s16le', rate=sr)
+    output_stream.layout = 'mono'
+
+    for frame in container.decode(audio=0):
+        frame.pts = None
+        resampled_frames = resampler.resample(frame)
+        if resampled_frames is not None:
+            for resampled_frame in resampled_frames:
+                for packet in output_stream.encode(resampled_frame):
+                    output_container.mux(packet)
+
+    for packet in output_stream.encode(None):
+        output_container.mux(packet)
+
+    output_container.close()
+    return resampled_file

whisper_live/vad.py

@@ -0,0 +1,155 @@
+# original: https://github.com/snakers4/silero-vad/blob/master/utils_vad.py
+
+import os
+import subprocess
+import torch
+import numpy as np
+import onnxruntime
+import warnings
+
+
+class VoiceActivityDetection():
+
+    def __init__(self, force_onnx_cpu=True):
+        path = self.download()
+
+        opts = onnxruntime.SessionOptions()
+        opts.log_severity_level = 3
+
+        opts.inter_op_num_threads = 1
+        opts.intra_op_num_threads = 1
+
+        if force_onnx_cpu and 'CPUExecutionProvider' in onnxruntime.get_available_providers():
+            self.session = onnxruntime.InferenceSession(path, providers=['CPUExecutionProvider'], sess_options=opts)
+        else:
+            self.session = onnxruntime.InferenceSession(path, providers=['CUDAExecutionProvider'], sess_options=opts)
+
+        self.reset_states()
+        self.sample_rates = [8000, 16000]
+
+    def _validate_input(self, x, sr: int):
+        if x.dim() == 1:
+            x = x.unsqueeze(0)
+        if x.dim() > 2:
+            raise ValueError(f"Too many dimensions for input audio chunk {x.dim()}")
+
+        if sr != 16000 and (sr % 16000 == 0):
+            step = sr // 16000
+            x = x[:, ::step]
+            sr = 16000
+
+        if sr not in self.sample_rates:
+            raise ValueError(f"Supported sampling rates: {self.sample_rates} (or multiply of 16000)")
+        if sr / x.shape[1] > 31.25:
+            raise ValueError("Input audio chunk is too short")
+
+        return x, sr
+
+    def reset_states(self, batch_size=1):
+        self._state = torch.zeros((2, batch_size, 128)).float()
+        self._context = torch.zeros(0)
+        self._last_sr = 0
+        self._last_batch_size = 0
+
+    def __call__(self, x, sr: int):
+
+        x, sr = self._validate_input(x, sr)
+        num_samples = 512 if sr == 16000 else 256
+
+        if x.shape[-1] != num_samples:
+            raise ValueError(f"Provided number of samples is {x.shape[-1]} (Supported values: 256 for 8000 sample rate, 512 for 16000)")
+
+        batch_size = x.shape[0]
+        context_size = 64 if sr == 16000 else 32
+
+        if not self._last_batch_size:
+            self.reset_states(batch_size)
+        if (self._last_sr) and (self._last_sr != sr):
+            self.reset_states(batch_size)
+        if (self._last_batch_size) and (self._last_batch_size != batch_size):
+            self.reset_states(batch_size)
+
+        if not len(self._context):
+            self._context = torch.zeros(batch_size, context_size)
+
+        x = torch.cat([self._context, x], dim=1)
+        if sr in [8000, 16000]:
+            ort_inputs = {'input': x.numpy(), 'state': self._state.numpy(), 'sr': np.array(sr, dtype='int64')}
+            ort_outs = self.session.run(None, ort_inputs)
+            out, state = ort_outs
+            self._state = torch.from_numpy(state)
+        else:
+            raise ValueError()
+
+        self._context = x[..., -context_size:]
+        self._last_sr = sr
+        self._last_batch_size = batch_size
+
+        out = torch.from_numpy(out)
+        return out
+
+    def audio_forward(self, x, sr: int):
+        outs = []
+        x, sr = self._validate_input(x, sr)
+        self.reset_states()
+        num_samples = 512 if sr == 16000 else 256
+
+        if x.shape[1] % num_samples:
+            pad_num = num_samples - (x.shape[1] % num_samples)
+            x = torch.nn.functional.pad(x, (0, pad_num), 'constant', value=0.0)
+
+        for i in range(0, x.shape[1], num_samples):
+            wavs_batch = x[:, i:i+num_samples]
+            out_chunk = self.__call__(wavs_batch, sr)
+            outs.append(out_chunk)
+
+        stacked = torch.cat(outs, dim=1)
+        return stacked.cpu()
+
+    @staticmethod
+    def download(model_url="https://github.com/snakers4/silero-vad/raw/v5.0/files/silero_vad.onnx"):
+        target_dir = os.path.expanduser("~/.cache/whisper-live/")
+
+        # Ensure the target directory exists
+        os.makedirs(target_dir, exist_ok=True)
+
+        # Define the target file path
+        model_filename = os.path.join(target_dir, "silero_vad.onnx")
+
+        # Check if the model file already exists
+        if not os.path.exists(model_filename):
+            # If it doesn't exist, download the model using wget
+            try:
+                subprocess.run(["wget", "-O", model_filename, model_url], check=True)
+            except subprocess.CalledProcessError:
+                print("Failed to download the model using wget.")
+        return model_filename
+
+
+class VoiceActivityDetector:
+    def __init__(self, threshold=0.5, frame_rate=16000):
+        """
+        Initializes the VoiceActivityDetector with a voice activity detection model and a threshold.
+
+        Args:
+            threshold (float, optional): The probability threshold for detecting voice activity. Defaults to 0.5.
+        """
+        self.model = VoiceActivityDetection()
+        self.threshold = threshold
+        self.frame_rate = frame_rate
+
+    def __call__(self, audio_frame):
+        """
+        Determines if the given audio frame contains speech by comparing the detected speech probability against
+        the threshold.
+
+        Args:
+            audio_frame (np.ndarray): The audio frame to be analyzed for voice activity. It is expected to be a
+                                      NumPy array of audio samples.
+
+        Returns:
+            bool: True if the speech probability exceeds the threshold, indicating the presence of voice activity;
+                  False otherwise.
+        """
+        speech_probs = self.model.audio_forward(torch.from_numpy(audio_frame.copy()), self.frame_rate)[0]
+        return torch.any(speech_probs > self.threshold).item()