fix(transcription): treat libsndfile's unknown frame count as no duration

Backport of #50842 to lts/1.108.x.
Cherry-picked from e6a0534785 (litellm_transcription_unknown_frame_count).

litellm/litellm_core_utils/audio_utils/utils.py

@@ -28,6 +28,13 @@ def normalized_audio_duration_seconds(value: object) -> float | None:
     return seconds if math.isfinite(seconds) and seconds >= 0 else None
 
 
+LIBSNDFILE_UNKNOWN_FRAME_COUNT: Final = 2**63 - 1
+
+
+def soundfile_duration_seconds(frames: int, samplerate: int) -> float | None:
+    return None if frames >= LIBSNDFILE_UNKNOWN_FRAME_COUNT else frames / samplerate
+
+
 @dataclass
 class ProcessedAudioFile:
     """
@@ -334,8 +341,7 @@ def calculate_request_duration(file: FileTypes) -> float | None:
         # Extract duration using soundfile
         file_object: Final = io.BytesIO(file_content)
         with sf.SoundFile(file_object) as audio:
-            duration: Final = len(audio) / audio.samplerate
-            return duration
+            return soundfile_duration_seconds(len(audio), audio.samplerate)
 
     except Exception:
         # Silently fail if duration extraction fails

litellm/proxy/pass_through_endpoints/llm_provider_handlers/transcribe_passthrough_logging_handler.py

@@ -29,6 +29,7 @@ from litellm.constants import (
     TRANSCRIBE_MEDIA_FETCH_ATTEMPTS,
     TRANSCRIBE_MEDIA_LAST_MODIFIED_TOLERANCE_SECONDS,
 )
+from litellm.litellm_core_utils.audio_utils.utils import soundfile_duration_seconds
 from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.litellm_core_utils.litellm_logging import (
@@ -513,7 +514,7 @@ async def write_media_within_limit(response: httpx.Response, media_file: IO[byte
 def media_file_seconds(path: Path) -> float | None:
     try:
         with soundfile.SoundFile(str(path)) as audio:
-            return len(audio) / audio.samplerate
+            return soundfile_duration_seconds(len(audio), audio.samplerate)
     except (RuntimeError, ValueError, OSError) as e:
         verbose_proxy_logger.warning("Transcribe media could not be decoded for its duration: %s", e)
         return None

tests/integration/cost_calculation/cases/openai.py

@@ -19,6 +19,7 @@ from integration.cost_calculation.cost_tracking_case import (
     RecountExpected,
     RecountRates,
     SseResponse,
+    UnsizedFlacUpload,
     WavUpload,
 )
 from pydantic import JsonValue
@@ -7743,6 +7744,18 @@ WHISPER_NEXT_TRANSCRIPTIONS_PER_SECOND: Final = CostTrackingTestCase(
 )
 
 
+WHISPER_NEXT_TRANSCRIPTIONS_UNSIZED_FLAC_NO_USAGE: Final = CostTrackingTestCase(
+    name="whisper-next-transcriptions-unsized-flac-no-usage",
+    covers="quota_management.spend_tracking.cost_matrix.logs_cost",
+    model="whisper-next",
+    endpoint="/v1/audio/transcriptions",
+    upload=UnsizedFlacUpload(kind="unsized_flac"),
+    request={"response_format": "json"},
+    response=JsonResponse(content_type="application/json", body={"text": "hello"}),
+    expected=ExactExpected(spend=0, input_cost=0, output_cost=0, prompt_tokens=0, completion_tokens=0),
+)
+
+
 # whisper-verbose-next
 WHISPER_VERBOSE_NEXT_TRANSCRIPTIONS_DURATION: Final = CostTrackingTestCase(
     name="whisper-verbose-next-transcriptions-duration",
@@ -11391,6 +11404,7 @@ CASES: Final[tuple[CostTrackingTestCase, ...]] = (
     GPT_5_6_UPSTREAM_500_ZERO_SPEND,
     GPT_5_6_UPSTREAM_429_ZERO_SPEND,
     WHISPER_NEXT_TRANSCRIPTIONS_PER_SECOND,
+    WHISPER_NEXT_TRANSCRIPTIONS_UNSIZED_FLAC_NO_USAGE,
     WHISPER_VERBOSE_NEXT_TRANSCRIPTIONS_DURATION,
     GPT_4O_TRANSCRIBE_NEXT_TRANSCRIPTIONS_TOKENS,
     TTS_NEXT_SPEECH_PER_CHARACTER,

tests/integration/cost_calculation/cost_tracking_case.py

@@ -168,7 +168,13 @@ class PngUpload(BaseModel):
     kind: Literal["png"]
 
 
-Upload: TypeAlias = Annotated[WavUpload | PngUpload, Field(discriminator="kind")]
+class UnsizedFlacUpload(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    kind: Literal["unsized_flac"]
+
+
+Upload: TypeAlias = Annotated[WavUpload | PngUpload | UnsizedFlacUpload, Field(discriminator="kind")]
 
 
 class JsonResponse(BaseModel):

tests/integration/cost_calculation/test_cost_tracking.py

@@ -72,6 +72,19 @@ def _wav_bytes(seconds: float) -> bytes:
     return output.getvalue()
 
 
+def _unsized_flac_bytes() -> bytes:
+    """A FLAC whose STREAMINFO total-samples field is 0 (unknown), as an encoder writing to a pipe leaves it."""
+    sample_rate: Final = 16000
+    bits_per_sample: Final = 16
+    streaminfo: Final = (
+        struct.pack(">HH", 4096, 4096)
+        + bytes(6)
+        + ((sample_rate << 44) | ((bits_per_sample - 1) << 36)).to_bytes(8, "big")
+        + bytes(16)
+    )
+    return b"fLaC" + b"\x80" + len(streaminfo).to_bytes(3, "big") + streaminfo
+
+
 def _png_bytes() -> bytes:
     def chunk(kind: bytes, payload: bytes) -> bytes:
         return (
@@ -97,6 +110,8 @@ def _multipart_request(gateway: Gateway, case: CostTrackingTestCase, model_name:
     }
     if case.upload.kind == "wav":
         files: Final = {"file": ("audio.wav", _wav_bytes(case.upload.seconds), "audio/wav")}
+    elif case.upload.kind == "unsized_flac":
+        files = {"file": ("audio.flac", _unsized_flac_bytes(), "audio/flac")}
     else:
         files = {"image": ("image.png", _png_bytes(), "image/png")}
     return gateway.request_multipart(case.endpoint, fields, files, key=key)

tests/unit/litellm_core_utils/test_audio_utils.py

@@ -4,7 +4,9 @@ Test the audio utils functionality in litellm_core_utils/audio_utils/utils.py
 
 import io
 import os
+import struct
 import tempfile
+from typing import Final
 from unittest.mock import mock_open, patch
 
 import pytest
@@ -277,6 +279,17 @@ class TestCalculateRequestDuration:
         ), "File position should be restored to original position"
 
 
+def _unsized_flac() -> bytes:
+    streaminfo: Final = (
+        struct.pack(">HH", 4096, 4096) + bytes(6) + ((16000 << 44) | (15 << 36)).to_bytes(8, "big") + bytes(16)
+    )
+    return b"fLaC" + b"\x80" + len(streaminfo).to_bytes(3, "big") + streaminfo
+
+
+def test_duration_is_unknown_for_a_flac_that_does_not_state_its_length() -> None:
+    assert calculate_request_duration(_unsized_flac()) is None
+
+
 class TestGetAudioFileContentHash:
     """Test the get_audio_file_content_hash function for cache key generation"""
 

tests/unit/proxy/pass_through_endpoints/llm_provider_handlers/test_transcribe_passthrough_logging_handler.py

@@ -1,6 +1,7 @@
 import asyncio
 import io
 import json
+import struct
 import wave
 from datetime import datetime
 from pathlib import Path
@@ -453,6 +454,15 @@ class TestMediaFileSeconds:
 
         assert media_file_seconds(media) == pytest.approx(1.5)
 
+    def test_flac_that_does_not_state_its_length_yields_no_duration(self, tmp_path: Path):
+        streaminfo = (
+            struct.pack(">HH", 4096, 4096) + bytes(6) + ((16000 << 44) | (15 << 36)).to_bytes(8, "big") + bytes(16)
+        )
+        media = tmp_path / "a.flac"
+        _ = media.write_bytes(b"fLaC" + b"\x80" + len(streaminfo).to_bytes(3, "big") + streaminfo)
+
+        assert media_file_seconds(media) is None
+
     def test_undecodable_media_yields_no_duration(self, tmp_path: Path):
         media = tmp_path / "a.wav"
         _ = media.write_bytes(b"not audio at all")