lattifai 0.4.3__py3-none-any.whl → 0.4.5__py3-none-any.whl

lattifai/bin/agent.py

@@ -302,6 +302,10 @@ if not check_dependencies():
 if __name__ == '__main__':
     import os
 
+    from dotenv import find_dotenv, load_dotenv
+
+    load_dotenv(find_dotenv(usecwd=True))
+
     asyncio.run(
         _run_youtube_workflow(
             # url='https://www.youtube.com/watch?v=7nv1snJRCEI',

lattifai/io/text_parser.py

@@ -8,12 +8,12 @@ SPEAKER_PATTERN = re.compile(r'((?:>>|>>|>|>).*?[:：])\s*(.*)')
 # Transcriber Output Example:
 # 26:19.919 --> 26:34.921
 # [SPEAKER_01]: 越来越多的科技巨头入...
-SPEAKER_LATTIFAI = re.compile(r'(^\[SPEAKER_.*?\]:)\s*(.*)')
+SPEAKER_LATTIFAI = re.compile(r'(^\[SPEAKER_.*?\][:：])\s*(.*)')
 
 # NISHTHA BHATIA: Hey, everyone.
 # DIETER: Oh, hey, Nishtha.
 # GEMINI: That might
-SPEAKER_PATTERN2 = re.compile(r'^([A-Z]{1,15}(?:\s+[A-Z]{1,15})?)[:：]\s*(.*)$')
+SPEAKER_PATTERN2 = re.compile(r'^([A-Z]{1,15}(?:\s+[A-Z]{1,15})?[:：])\s*(.*)$')
 
 
 def parse_speaker_text(line) -> Tuple[Optional[str], str]:

lattifai/tokenizer/tokenizer.py

@@ -148,7 +148,7 @@ class LatticeTokenizer:
         oov_words = []
         for text in texts:
             words = text.lower().replace('-', ' ').replace('—', ' ').replace('–', ' ').split()
-            oovs = [w for w in words if w not in self.words]
+            oovs = [w.strip(PUNCTUATION) for w in words if w not in self.words]
             if oovs:
                 oov_words.extend([w for w in oovs if (w not in self.words and len(w) <= MAXIMUM_WORD_LENGTH)])
 

lattifai/workers/lattice1_alpha.py

@@ -1,25 +1,70 @@
 import json
 import time
 from collections import defaultdict
-from typing import Any, BinaryIO, Dict, Tuple, Union
+from typing import Any, BinaryIO, Dict, Iterable, Optional, Tuple, Union
 
 import numpy as np
 import onnxruntime as ort
-import resampy
 import soundfile as sf
 import torch
 from lhotse import FbankConfig
-from lhotse.audio import read_audio
+from lhotse.augmentation import get_or_create_resampler
 from lhotse.features.kaldi.layers import Wav2LogFilterBank
 from lhotse.utils import Pathlike
 
-from lattifai.errors import (
-    AlignmentError,
-    AudioFormatError,
-    AudioLoadError,
-    DependencyError,
-    ModelLoadError,
-)
+from lattifai.errors import AlignmentError, AudioFormatError, AudioLoadError, DependencyError, ModelLoadError
+
+ChannelSelectorType = Union[int, Iterable[int], str]
+
+
+def resample_audio(
+    audio_sr: Tuple[torch.Tensor, int],
+    sampling_rate: int,
+    device: Optional[str],
+    channel_selector: Optional[ChannelSelectorType] = 'average',
+) -> torch.Tensor:
+    """
+    return:
+        (1, T)
+    """
+    audio, sr = audio_sr
+
+    if channel_selector is None:
+        # keep the original multi-channel signal
+        tensor = audio
+    elif isinstance(channel_selector, int):
+        assert audio.shape[0] >= channel_selector, f'Invalid channel: {channel_selector}'
+        tensor = audio[channel_selector : channel_selector + 1].clone()
+        del audio
+    elif isinstance(channel_selector, str):
+        assert channel_selector == 'average'
+        tensor = torch.mean(audio.to(device), dim=0, keepdim=True)
+        del audio
+    else:
+        assert isinstance(channel_selector, Iterable)
+        num_channels = audio.shape[0]
+        print(f'Selecting channels {channel_selector} from the signal with {num_channels} channels.')
+        assert isinstance(channel_selector, Iterable)
+        if max(channel_selector) >= num_channels:
+            raise ValueError(
+                f'Cannot select channel subset {channel_selector} from a signal with {num_channels} channels.'
+            )
+        tensor = audio[channel_selector]
+
+    tensor = tensor.to(device)
+    if sr != sampling_rate:
+        resampler = get_or_create_resampler(sr, sampling_rate).to(device=device)
+        length = tensor.size(-1)
+        chunk_size = sampling_rate * 3600
+        if length > chunk_size:
+            resampled_chunks = []
+            for i in range(0, length, chunk_size):
+                resampled_chunks.append(resampler(tensor[..., i : i + chunk_size]))
+            tensor = torch.cat(resampled_chunks, dim=-1)
+        else:
+            tensor = resampler(tensor)
+
+    return tensor
 
 
 class Lattice1AlphaWorker:
@@ -48,7 +93,7 @@ class Lattice1AlphaWorker:
             self.acoustic_ort = ort.InferenceSession(
                 f'{model_path}/acoustic_opt.onnx',
                 sess_options,
-                providers=providers + ['CoreMLExecutionProvider', 'CPUExecutionProvider'],
+                providers=providers + ['CPUExecutionProvider', 'CoreMLExecutionProvider'],
             )
         except Exception as e:
             raise ModelLoadError(f'acoustic model from {model_path}', original_error=e)
@@ -93,17 +138,13 @@ class Lattice1AlphaWorker:
         self.timings['emission'] += time.time() - _start
         return emission  # (1, T, vocab_size) torch
 
-    def load_audio(self, audio: Union[Pathlike, BinaryIO]) -> Tuple[torch.Tensor, int]:
+    def load_audio(
+        self, audio: Union[Pathlike, BinaryIO], channel_selector: Optional[ChannelSelectorType] = 'average'
+    ) -> Tuple[torch.Tensor, int]:
         # load audio
         try:
-            waveform, sample_rate = read_audio(audio)  # numpy array
-            if len(waveform.shape) == 1:
-                waveform = waveform.reshape([1, -1])  # (1, L)
-            else:  # make sure channel first
-                if waveform.shape[0] > waveform.shape[1]:
-                    waveform = waveform.transpose(0, 1)
-                # average multiple channels
-                waveform = np.mean(waveform, axis=0, keepdims=True)  # (1, L)
+            waveform, sample_rate = sf.read(audio, always_2d=True, dtype='float32')  # numpy array
+            waveform = waveform.T  # (channels, samples)
         except Exception as primary_error:
             # Fallback to PyAV for formats not supported by soundfile
             try:
@@ -141,23 +182,16 @@ class Lattice1AlphaWorker:
 
                 # Concatenate all frames
                 waveform = np.concatenate(frames, axis=1)
-                # Average multiple channels to mono
-                if waveform.shape[0] > 1:
-                    waveform = np.mean(waveform, axis=0, keepdims=True)
-
                 sample_rate = audio_stream.codec_context.sample_rate
             except Exception as e:
                 raise AudioLoadError(str(audio), original_error=e)
 
-        try:
-            if sample_rate != self.config['sample_rate']:
-                waveform = resampy.resample(waveform, sample_rate, self.config['sample_rate'], axis=1)
-        except Exception:
-            raise AudioFormatError(
-                str(audio), f'Failed to resample from {sample_rate}Hz to {self.config["sample_rate"]}Hz'
-            )
-
-        return torch.from_numpy(waveform).to(self.device)  # (1, L)
+        return resample_audio(
+            (torch.from_numpy(waveform), sample_rate),
+            self.config.get('sampling_rate', 16000),
+            device=self.device.type,
+            channel_selector=channel_selector,
+        )
 
     def alignment(
         self, audio: Union[Union[Pathlike, BinaryIO], torch.tensor], lattice_graph: Tuple[str, int, float]
@@ -231,9 +265,9 @@ class Lattice1AlphaWorker:
                 emission.to(device) * acoustic_scale,
                 decoding_graph.to(device),
                 torch.tensor([emission.shape[1]], dtype=torch.int32),
-                search_beam=100,
-                output_beam=40,
-                min_active_states=200,
+                search_beam=200,
+                output_beam=80,
+                min_active_states=400,
                 max_active_states=10000,
                 subsampling_factor=1,
                 reject_low_confidence=False,

lattifai/workflows/youtube.py

@@ -708,7 +708,7 @@ class YouTubeSubtitleAgent(WorkflowAgent):
 
         # If subtitle was already downloaded in step 1 and user selected it, use it directly
         if downloaded_subtitle_path and downloaded_subtitle_path != 'gemini':
-            self.logger.info(f'📥 Using subtitle downloaded in previous step: {downloaded_subtitle_path}')
+            self.logger.info(f'📥 Using subtitle: {downloaded_subtitle_path}')
             return {'subtitle_path': downloaded_subtitle_path}
 
         # Check for existing subtitles if subtitle was not downloaded yet

{lattifai-0.4.3.dist-info → lattifai-0.4.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lattifai
-Version: 0.4.3
+Version: 0.4.5
 Summary: Lattifai Python SDK: Seamless Integration with Lattifai's Speech and Video AI Services
 Author-email: Lattifai Technologies <tech@lattifai.com>
 Maintainer-email: Lattice <tech@lattifai.com>
@@ -57,7 +57,6 @@ Requires-Dist: pysubs2
 Requires-Dist: praatio
 Requires-Dist: tgt
 Requires-Dist: onnxruntime
-Requires-Dist: resampy
 Requires-Dist: g2p-phonemizer==0.1.1
 Requires-Dist: wtpsplit>=2.1.6
 Requires-Dist: av

{lattifai-0.4.3.dist-info → lattifai-0.4.5.dist-info}/RECORD

@@ -4,7 +4,7 @@ lattifai/client.py,sha256=YvK25fcXwKREYCkq_TUBdEZh7I9RNEwCbgW4qUha2ho,13236
 lattifai/errors.py,sha256=5i_acoly1g-TLAID8QnhzQshwOXfgLL55mHsdwzlNGA,10814
 lattifai/utils.py,sha256=CzVwNc08u8lm7XavCMJskXHfni0xsZ-EgpcMkRywVm8,4736
 lattifai/bin/__init__.py,sha256=QWbmVUbzqwPQNeOV_g-bOsz53w37v-tbZ3uFrSj-7Mg,90
-lattifai/bin/agent.py,sha256=jc7qSuVV2_EkToEu2qYodXgGoTup_na6IgP25kFmfwk,9734
+lattifai/bin/agent.py,sha256=8nRh0GC1M4__-qKQtxZspcyNnUm8DOVCWQf3sRblEOI,9826
 lattifai/bin/align.py,sha256=JJHk5uzmYGYhpA3ricHdmBLzJWC2aRyF0k4ANnap50w,8151
 lattifai/bin/cli_base.py,sha256=gvPUi9Z0eGwBJ8Es5xq1z00YzFPlocYiR2zpL7ekyhw,626
 lattifai/bin/subtitle.py,sha256=UZMPh71O2X1UwbfZ9VWlhzxkz78viz8KWwoVsDpewK0,6577
@@ -13,27 +13,27 @@ lattifai/io/gemini_reader.py,sha256=WDZA93MSrUAsa5j-ZDXLdPXzEIoREymEy-rMAED_6f4,
 lattifai/io/gemini_writer.py,sha256=rlXO9zx6kQhqTi9K9izE69-8S-2GPOIiJHPwZyebpiM,6515
 lattifai/io/reader.py,sha256=h4T8dveLHXqSonma0J50iyjqkxH26tujeoPbnLx05nA,3333
 lattifai/io/supervision.py,sha256=iBDRiDJ0hddo__SoEZau2cdEIBFnXZNLgSWFjtJd-lM,871
-lattifai/io/text_parser.py,sha256=LQHgcEYXaSdhwUo9rP6P_31Z6RMv_BTP1YSKzXji4bk,2386
+lattifai/io/text_parser.py,sha256=5rO2qgFLfWIcoU0K-FGBgs2qG7r6Uz2J2LklGTgSUyA,2391
 lattifai/io/utils.py,sha256=4drRwcM1n7AYhdJcF51EZxMTy_Ut_1GKtdWpRhPuVmg,686
 lattifai/io/writer.py,sha256=8n9ZBuXuVOCFwzr1hqrnXpZ-fARTsepebwjKgRuueWE,3872
 lattifai/tokenizer/__init__.py,sha256=y-FyfO7tLga9b46pkCC6jdSBKOFZS-jFfHcqUieGEyU,120
 lattifai/tokenizer/phonemizer.py,sha256=SfRi1KIMpmaao6OVmR1h_I_3QU-vrE6D5bh72Afg5XM,1759
-lattifai/tokenizer/tokenizer.py,sha256=6XlHehhwahVWgUMj79LqMs13xaRTa17beOZqico4vLE,19186
+lattifai/tokenizer/tokenizer.py,sha256=uLtGeT6ad_vPuoXNw8Rz43bAdKQWheI_tfHAV47FyqQ,19205
 lattifai/workers/__init__.py,sha256=s6YfkIq4FDIAzY9sPjRpXnJfszj2repqnMTqydRM5Zw,83
-lattifai/workers/lattice1_alpha.py,sha256=1lCq0-bgWMXvYslAbCTFgHC0p6UWPto1y0wkTw9WrmQ,10177
+lattifai/workers/lattice1_alpha.py,sha256=4FX91eTmcTDZt78NrMY8EgkwlwBKlU4Qw39qcQeqiTg,11469
 lattifai/workflows/__init__.py,sha256=mTOdwQQs2YY1s0JsVGsATb2TWPhpNo7bRiAAJW92740,830
 lattifai/workflows/agents.py,sha256=ZnxyEb-Li8gQw3Z7V3K7jjnT7sQAtG1uCFuXGX77IcE,227
 lattifai/workflows/base.py,sha256=ZSePq2O2hB59K5CMbk_iGiOM9FSHOVQdG3t8Oyz9gwE,6264
 lattifai/workflows/file_manager.py,sha256=5km59cB4s9PcJgMMqZ6YqRY8sTqvTzd1qyQ2T5bHlfI,31985
 lattifai/workflows/gemini.py,sha256=kpp3GiHyui8RZLWRQMx7vojBscCi2lZSxrjM1VHo_mc,6144
-lattifai/workflows/youtube.py,sha256=y1ez32G40CtHAo7Kz7J_DO96ljOwW-3SP9Sm2uPe5hY,38785
+lattifai/workflows/youtube.py,sha256=tRy9LV5mwgQfQ3jfL_aiyB2catDHqCS2Hmy9Lk3-VGs,38757
 lattifai/workflows/prompts/README.md,sha256=X49KWSQVdjWxxWUp4R2w3ZqKrAOi6_kDNHh1hMaQ4PE,694
 lattifai/workflows/prompts/__init__.py,sha256=i3kMT5sg_W9LbPcda0xmZWLg0tPjXGVI3iKtHrBng3o,1351
 lattifai/workflows/prompts/gemini/README.md,sha256=rt7f7yDGtaobKBo95LG3u56mqa3ABOXQd0UVgJYtYuo,781
 lattifai/workflows/prompts/gemini/transcription_gem.txt,sha256=cljzZ--BDgnnKzqVCakr-fTp2Xk38UOsUquvruNX-LU,4600
-lattifai-0.4.3.dist-info/licenses/LICENSE,sha256=LNuoH5jpXXNKgjQ3XLwztFq8D3O7kZI-LSg81o4ym2M,1065
-lattifai-0.4.3.dist-info/METADATA,sha256=-WztCYs961omWHWJMd1PcM6KLjQgbGYzYl72xGqMYdE,26710
-lattifai-0.4.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-lattifai-0.4.3.dist-info/entry_points.txt,sha256=fCgo8-LKA_9C7_jmEGsZPJko0woXHtEh0iRbpO7PYzI,69
-lattifai-0.4.3.dist-info/top_level.txt,sha256=tHSoXF26r-IGfbIP_JoYATqbmf14h5NrnNJGH4j5reI,9
-lattifai-0.4.3.dist-info/RECORD,,
+lattifai-0.4.5.dist-info/licenses/LICENSE,sha256=LNuoH5jpXXNKgjQ3XLwztFq8D3O7kZI-LSg81o4ym2M,1065
+lattifai-0.4.5.dist-info/METADATA,sha256=e0ojxc-4xgpgpWZC4wdXlTySQYliuLCa-MGG1F80U6E,26687
+lattifai-0.4.5.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+lattifai-0.4.5.dist-info/entry_points.txt,sha256=fCgo8-LKA_9C7_jmEGsZPJko0woXHtEh0iRbpO7PYzI,69
+lattifai-0.4.5.dist-info/top_level.txt,sha256=tHSoXF26r-IGfbIP_JoYATqbmf14h5NrnNJGH4j5reI,9
+lattifai-0.4.5.dist-info/RECORD,,