`AudioData.get_wav_data()` only processes partial audio data, causing incomplete transcriptions

[ftnext]

Bug Description

When using the SpeechRecognition library with OpenAI Whisper API, only the first few seconds of audio files are transcribed, regardless of the actual file duration or size.

Steps to Reproduce

1. Use a WAV audio file longer than ~30 seconds
2. Run transcription using: python -m speech_recognition.recognizers.whisper_api.openai --model gpt-4o-transcribe audio_file.wav
3. Observe that only the first portion is transcribed

% uv run --python 3.12 --with 'SpeechRecognition[openai]==3.14.2' -- python -m speech_recognition.recognizers.whisper_api.openai -l ja long_audio.wav

Here is long_audio.wav example:
https://notebooklm.google.com/notebook/e7297b2e-e363-4e77-bff3-8d71e104d5a2

Expected Behavior

The entire audio file should be transcribed.

Actual Behavior

Only the first few seconds are transcribed (e.g., 18 characters from a 7.6-minute file).

Root Cause Analysis

The issue appears to be in AudioData.get_wav_data() method. When processing audio files, the method only converts a small portion of the audio data:

• Original file: 21.89 MB, 456 seconds
• WAV conversion result: 0.08 MB (abnormally small)
• This suggests only ~2-3 seconds of audio are being processed

Evidence

Testing the same audio file directly with OpenAI Python SDK works perfectly:

• SpeechRecognition library: 18 characters transcribed
• Direct OpenAI API: 2,829 characters (complete transcription)

Environment

• SpeechRecognition version: 3.14.2
• Python: 3.12
• Audio format: 24kHz, 16-bit, mono PCM WAV
• Models tested: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe (all show same issue)

Workaround

Use OpenAI Python SDK directly instead of SpeechRecognition library.

[ftnext]

Devin investigated! (many thanks)

Root Cause:
The bug is in the interaction between Recognizer.record() method and AudioFileStream.read() in /speech_recognition/__init__.py:

• Line 353: buffer = source.stream.read(source.CHUNK) always reads in chunks of 4096 frames
• Line 299: AudioFileStream.read() correctly supports reading all frames when size=-1, but record() never uses this
• The chunked reading loop terminates prematurely, capturing only the first few seconds of audio files

Technical Details:

• source.CHUNK = 4096 (line 280)
• For 24kHz audio: 4096/24000 ≈ 0.17 seconds per chunk
• Issue reports ~2-3 seconds processed = ~15-18 chunks before premature termination
• This explains why 21.89 MB, 456-second file → 0.08 MB WAV output

Historical Context:
The chunked reading behavior was present since the initial 2014 release and was designed for microphone input (real-time processing), not file input. The original AudioFileStream.read() implementation correctly supports reading entire files when size=-1.

Potential Fix:
Modify record() method to detect file sources and read all data at once when no duration is specified:

if isinstance(source, AudioFile) and duration is None and offset is None:
    buffer = source.stream.read(-1)  # Read all frames
    return AudioData(buffer, source.SAMPLE_RATE, source.SAMPLE_WIDTH)

This would restore the intended behavior for file processing while preserving chunked reading for microphone input.