Panniantong/Agent-Reach · error · TranscribeError

audio chunks total {total_chunk_bytes} bytes; safety limit i

Error message

audio chunks total {total_chunk_bytes} bytes; safety limit is {limit_mib:g} MiB

What it means

Raised when the sum of all chunk sizes exceeds MAX_TOTAL_CHUNK_BYTES (96 MiB) after per-chunk size validation passes. Even when each chunk is under the 24 MiB Whisper limit, the aggregate upload budget (cost/time guard) is capped at 96 MiB per transcribe() call. It fires before any upload, so no API spend has occurred.

Source

Thrown at agent_reach/transcribe.py:475

    if compressed.stat().st_size <= SIZE_LIMIT_BYTES:
        chunks = [compressed]
    else:
        chunks = chunk_audio(compressed, work_dir)

    if len(chunks) > MAX_CHUNKS:
        max_minutes = MAX_CHUNKS * CHUNK_SECONDS // 60
        raise TranscribeError(
            f"audio produced {len(chunks)} chunks; safety limit is "
            f"{MAX_CHUNKS} (~{max_minutes} minutes)"
        )
    chunk_sizes = [
        _require_size_at_most(chunk, SIZE_LIMIT_BYTES, f"chunk {chunk.name}")
        for chunk in chunks
    ]
    total_chunk_bytes = sum(chunk_sizes)
    if total_chunk_bytes > MAX_TOTAL_CHUNK_BYTES:
        limit_mib = MAX_TOTAL_CHUNK_BYTES / (1024 * 1024)
        raise TranscribeError(
            f"audio chunks total {total_chunk_bytes} bytes; "
            f"safety limit is {limit_mib:g} MiB"
        )

    pieces: List[str] = []
    for chunk in chunks:
        text = _transcribe_with_fallback(chunk, order, cfg)
        pieces.append(text.strip())
    return "\n".join(p for p in pieces if p)


def _transcribe_with_fallback(chunk: Path, order: List[str], config: Config) -> str:
    """Try each provider in order; return first success or raise the last error."""
    last_err: Optional[Exception] = None
    for p in order:
        if not _provider_key(p, config):
            # Skip silently — caller already validated at least one is configured.
            continue

View on GitHub (pinned to 93ae1d18c3)

Solutions

Reduce source length: trim with yt-dlp --download-sections or ffmpeg and transcribe parts separately
Re-encode the source at a lower bitrate (e.g. ffmpeg -b:a 48k mono flac/opus) before calling transcribe()
Check that ffmpeg compression is actually succeeding in your environment (doctor verifies ffmpeg presence)

Example fix

# before
text = transcribe("huge_recording.wav")  # chunks total > 96 MiB

# after — pre-compress to speech-friendly bitrate
# ffmpeg -i huge_recording.wav -ac 1 -b:a 32k small.opus
text = transcribe("small.opus")

Defensive patterns

Strategy: validation

Validate before calling

from agent_reach.transcribe import MAX_TOTAL_CHUNK_BYTES  # 96 MiB

def compressed_size_ok(path: str) -> bool:
    import os
    return os.path.getsize(path) <= MAX_TOTAL_CHUNK_BYTES

Try / catch

from agent_reach.transcribe import transcribe, TranscribeError
try:
    text = transcribe(src)
except TranscribeError as e:
    if "chunks total" in str(e):
        src = recompress_mono_low_bitrate(src)  # ffmpeg -ac 1 -b:a 32k
        text = transcribe(src)
    else:
        raise

Prevention

Pre-encode speech audio to mono, low-bitrate opus/flac before transcribe()
Remember both caps: per-chunk 24 MiB, aggregate 96 MiB, count 24 chunks
Check compressed output size, not source size — compression ratio decides

When it happens

Trigger: Calling transcribe() on audio where chunk_audio produces e.g. 5 chunks of ~20 MiB each (100 MiB total) — each passes the per-chunk SIZE_LIMIT_BYTES check but the sum exceeds MAX_TOTAL_CHUNK_BYTES. Typical for high-bitheadroom audio that compresses poorly.

Common situations: High-bitrate music-heavy or noise-heavy sources (podcasts with music beds, concert recordings); WAV/FLAC sources converted at conservative compression settings; long audio near but under the 4-hour chunk-count cap.

Related errors

AI-assisted analysis of Panniantong/Agent-Reach@93ae1d18c3 (2026-08-14). Data as JSON: /api/errors/b66aac71a73753c7. Report an issue: GitHub.