Summary

This PR implements proper timestamp handling rules in the Whisper decoder to ensure accurate and well-formed timestamp generation during transcription, according to: https://github.com/openai/whisper/blob/e8622f9afc4eba139bf796c210f5c01081000472/whisper/decoding.py#L439

New Example Output

cargo run --example whisper --release --features="symphonia,metal"
    Finished `release` profile [optimized] target(s) in 0.44s
     Running `target/release/examples/whisper`
No audio file submitted: Downloading https://huggingface.co/datasets/Narsil/candle_demo/blob/main/samples_jfk.wav
pcm data loaded 176000
loaded mel: [1, 80, 3000]
0.0s -- 30.0s
  0.0s-1.8s: 
  1.8s-10.4s:  and so my fellow Americans ask not what your country can do for you ask what you can do for your country

Changes Made

Core Timestamp Rules Implementation

• Added apply_timestamp_rules() method that enforces OpenAI Whisper's timestamp constraints
• Implemented 4 key timestamp rules:
  1. Timestamp Pairing: Timestamps must come in pairs, except directly before EOT
  2. Non-decreasing Order: Timestamps cannot decrease in value
  3. Forced Initial Timestamp: Suppress non-timestamp tokens at the beginning of transcription
  4. Probability-based Preference: When timestamp probability sum exceeds any individual token, only consider timestamps

Configuration Enhancements

• Added max_initial_timestamp_index parameter to limit the maximum initial timestamp value
• Changed timestamps default to true (was previously disabled by default)
• Updated decoder constructor to accept the new timestamp constraint parameter

Code Structure Improvements

• Refactored decoder methods to use self.model consistently instead of mutable borrows
• Enhanced timestamp token detection and validation logic
• Added comprehensive masking system for applying multiple timestamp constraints

Benefits

✅ More accurate timestamp generation following OpenAI Whisper specifications
✅ Prevents malformed timestamp sequences
✅ Configurable initial timestamp constraints for better control
✅ Improved transcription quality with proper temporal alignment

Testing

The implementation follows the timestamp rules from the original OpenAI Whisper codebase, ensuring compatibility and correctness.