Add AudioData.split() for chunking large audio data

[ftnext]

Summary

Adds AudioData.split(max_bytes, *, silence_aware=False) so users can chunk oversized recordings into pieces that fit within API upload limits (e.g., the 25 MB cap on OpenAI's Whisper transcription endpoint).

• silence_aware=False (default): mechanical fixed-time split. No optional dependency. Each chunk's WAV-serialized size is <= max_bytes.
• silence_aware=True: snaps chunk boundaries to nearby silences via librosa.effects.split, looking only backward from the size-derived target so the cap is preserved. Requires the new audio-split extra (librosa, numpy). Lazy / numba-cache initialization failures are translated into SetupError.

Sample-aligned frame_data is required (split() raises ValueError on unaligned input) so the byte budget is a hard ceiling in both modes.

Usage

import speech_recognition as sr

audio = sr.AudioData.from_file("large.wav")
r = sr.Recognizer()

# 24 MB target leaves a small buffer under Whisper's 25 MB request limit.
chunks = audio.split(max_bytes=24 * 1024 * 1024, silence_aware=True)
text = " ".join(r.recognize_openai(c, model="whisper-1") for c in chunks)

Test plan

• 17 new tests in tests/test_audio.py:
  • Strict size cap in both modes
  • Sample-width coverage (8 / 16 / 24 / 32-bit) and explicit little-endian decoding (host-byte-order independence)
  • SetupError on missing librosa, on lazy import failures, and on call-time numba / runtime errors
  • Silence-aware boundary snaps to speech end within the look-back window
  • ValueError on unaligned frame_data and on too-small max_bytes
• CI coverage:
  • New audio-split entry in the extra-contracts matrix runs pytest tests/test_audio.py against .[dev,audio-split].
  • audio-split added to the Ubuntu all-extras install spec (skipped on 3.14 to match whisper-local).

🤖 Generated with Claude Code