Go Binding Does Not Support Incremental Processing

[joshmux]

Summary

In the Go bindings, calling Process() more than once results in no segments being returned for all subsequent calls. After the first successful Process() → NextSegment() cycle, any additional Process() invocation yields empty output.

This makes it impossible to perform incremental or streaming transcription.

Expected Behavior

• Each call to Process() with new audio should produce new segments.
• NextSegment() should return those new segments, and ideally preserve previously produced ones.
• This matches how the core whisper.cpp context behaves when processing audio in multiple chunks.

Actual Behavior

• The first call to Process() works correctly and NextSegment() returns segments.
• Future calls to Process() produce zero segments, even when valid new audio is passed.
• NextSegment() repeatedly returns no results for all subsequent Process() calls.

Why This Matters

This fully breaks incremental and streaming use cases, including:

• Real-time / chunked audio processing
• Live transcription
• Processing long audio without loading the full file at once
• Updating transcripts as new audio arrives

Users would expect the Go binding to match the C++ library’s ability to handle multiple processing passes.

Steps to Reproduce

1. Initialize a whisper.Context.
2. Call Process() with audio chunk A.
3. Call NextSegment() → segments from chunk A are returned as expected.
4. Call Process() again with audio chunk B.
5. Call NextSegment() → no segments are returned (output is empty).

Additional Notes

• This appears related to internal state handling within the Go binding.
• A related fix attempt is in PR Allow NextSegment() to be called across multiple Process() calls in the Go binding #3503.
• The issue may stem from the binding not resetting or managing internal counters needed for multi-pass processing.