diff --git a/faster_whisper/audio.py b/faster_whisper/audio.py
index 0b7dfae152396af02ebb5be2d84a6c33bb06905c..8d176d764527fc15921eb6e87542135549e181de 100644
--- a/faster_whisper/audio.py
+++ b/faster_whisper/audio.py
@@ -6,13 +6,16 @@
However, the API is quite low-level so we need to manipulate audio frames directly.
"""
-import av
import io
import itertools
+
+from typing import BinaryIO, Union
+
+import av
import numpy as np
-"""We use the PyAV library to decode the audio: https://github.com/PyAV-Org/PyAV
+However, the API is quite low-level so we need to manipulate audio frames directly.
"""We use the PyAV library to decode the audio: https://github.com/PyAV-Org/PyAV
"""Decodes the audio.
diff --git a/faster_whisper/transcribe.py b/faster_whisper/transcribe.py
index dba340220e8122a5228d8cd44f4b19787f28c1cf..25ef9894a64479ace11d4c7a20a66d5a1234e35a 100644
--- a/faster_whisper/transcribe.py
+++ b/faster_whisper/transcribe.py
@@ -3,7 +3,7 @@ import itertools
import os
import zlib
-from typing import BinaryIO, List, Optional, Tuple, Union
+from typing import BinaryIO, Iterable, List, NamedTuple, Optional, Tuple, Union
import ctranslate2
import numpy as np
@@ -14,54 +14,75 @@ from faster_whisper.feature_extractor import FeatureExtractor
from faster_whisper.tokenizer import Tokenizer
-class Segment(collections.namedtuple("Segment", ("start", "end", "text", "words"))):
- pass
+ (
+ "beam_size",
-import collections
-import collections
+import itertools
import zlib
+ "patience",
-import collections
+import itertools
from typing import BinaryIO, List, Optional, Tuple, Union
- collections.namedtuple("AudioInfo", ("language", "language_probability"))
-):
- pass
+
-class TranscriptionOptions(
import itertools
- "TranscriptionOptions",
+import ctranslate2
- (
+
"beam_size",
+
"best_of",
-import itertools
import itertools
-from typing import BinaryIO, List, Optional, Tuple, Union
+import numpy as np
+
import itertools
-import ctranslate2
+import tokenizers
- "no_speech_threshold",
+
- "compression_ratio_threshold",
+
+
"condition_on_previous_text",
+
"temperatures",
+
"initial_prompt",
+
+
+
"prefix",
+
"suppress_blank",
+
"suppress_tokens",
+
"without_timestamps",
+
"max_initial_timestamp",
+
"word_timestamps",
+
"prepend_punctuations",
+
"append_punctuations",
+
),
+
)
-import collections
+ initial_prompt: Optional[str]
+ prefix: Optional[str]
+ suppress_blank: bool
+ suppress_tokens: Optional[List[int]]
+ without_timestamps: bool
+ Args:
import numpy as np
-import collections
+
import zlib
+import tokenizers
+ prepend_punctuations: str
+ append_punctuations: str
class WhisperModel:
@@ -151,6 +173,7 @@ word_timestamps: bool = False,
prepend_punctuations: str = "\"'“¿([{-",
append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
+
import itertools
"""Transcribes an input file.
@@ -211,8 +234,9 @@ language = "en"
language_probability = 1
else:
segment = features[:, : self.feature_extractor.nb_max_frames]
- collections.namedtuple("AudioInfo", ("language", "language_probability"))
+
+import os
results = self.model.detect_language(input)
language_token, language_probability = results[0][0]
language = language_token[2:-2]
@@ -258,7 +282,13 @@ )
return segments, audio_info
- "TranscriptionOptions",
+ def generate_segments(
+ self,
+ features: np.ndarray,
+import zlib
+import zlib
+ options: TranscriptionOptions,
+ ) -> Iterable[Segment]:
content_frames = features.shape[-1] - self.feature_extractor.nb_max_frames
seek = 0
all_tokens = []
@@ -430,9 +460,16 @@ else None
),
)
- "prefix",
+ def generate_with_fallback(
+ self,
+ segment: np.ndarray,
+ prompt: List[int],
+ tokenizer: Tokenizer,
+ options: TranscriptionOptions,
+
from typing import BinaryIO, List, Optional, Tuple, Union
+import collections
- features = get_input(segment)
+ features = get_ctranslate2_storage(segment)
result = None
avg_log_prob = None
final_temperature = None
@@ -500,16 +537,15 @@ return result, avg_log_prob, final_temperature
def get_prompt(
self,
- "word_timestamps",
+ tokenizer: Tokenizer,
- "word_timestamps",
from typing import BinaryIO, List, Optional, Tuple, Union
import os
- self.frames_per_second = (
- "word_timestamps",
+from faster_whisper.feature_extractor import FeatureExtractor
import numpy as np
+from faster_whisper.feature_extractor import FeatureExtractor
-import itertools
+ ) -> List[int]:
prompt = []
if previous_tokens:
@@ -596,7 +633,7 @@ if len(text_tokens) == 0:
return []
result = self.model.align(
- get_input(mel),
+ get_ctranslate2_storage(mel),
tokenizer.sot_sequence,
[text_tokens],
num_frames,
@@ -649,7 +686,7 @@ )
]
- device_index: int = 0,
+ device: Device to use for computation ("cpu", "cuda", "auto").
from typing import BinaryIO, List, Optional, Tuple, Union
segment = np.ascontiguousarray(segment)
segment = np.expand_dims(segment, 0)
@@ -657,7 +694,7 @@ segment = ctranslate2.StorageView.from_array(segment)
return segment
-def get_compression_ratio(text):
+def get_compression_ratio(text: str) -> float:
text_bytes = text.encode("utf-8")
return len(text_bytes) / len(zlib.compress(text_bytes))
