feat: add FixedChunked transcriber

[humblemuzzu]

Reworked from #71 per your feedback — chunking now lives in the Transcriber layer, not inside the model.

What

FixedChunked — a new Transcriber implementation that splits audio into fixed-duration chunks with configurable overlap. Sits alongside VadChunked and EnergyAdaptiveChunked.

No VAD model, no energy analysis. The simplest chunking strategy for models with hard sequence-length limits.

Why overlap

EnergyAdaptiveChunked with search_window=0 gives you fixed-duration chunks, but with hard cuts. FixedChunked keeps a configurable tail from each chunk in the buffer so the next chunk starts with shared audio context. This prevents the model from seeing a hard cut mid-word at chunk boundaries.

Usage

use transcribe_rs::transcriber::{FixedChunked, FixedChunkedConfig, Transcriber};

let config = FixedChunkedConfig::default(); // 30s chunks, 1s overlap
let mut chunker = FixedChunked::new(config, TranscribeOptions::default());
let result = chunker.transcribe(&mut model, &samples)?;

Config

pub struct FixedChunkedConfig {
    pub chunk_duration_secs: f32,  // default 30.0
    pub overlap_secs: f32,         // default 1.0 (0.0 for hard cuts)
    pub padding_secs: f32,         // default 0.0
    pub min_chunk_secs: f32,       // default 0.0
    pub merge_separator: String,   // default " "
}

What changed

• New file: src/transcriber/fixed_chunked.rs (413 lines)
• src/transcriber/mod.rs: +3 lines (register, re-export, doc)

Uses existing transcribe_padded() and merge_sequential_with_separator(). No new dependencies. Parakeet model internals untouched.

Tests

12 unit tests using MockModel/FailOnNthModel, same patterns as the existing transcriber tests:

• Splitting at chunk duration
• Overlap retains tail correctly
• Remainder handled in finish()
• min_chunk_secs skips short remainders
• Timestamps correct with and without overlap
• Timestamp clamping with padding
• Empty input
• Object safety (Box<dyn Transcriber>)
• Reusable after error
• Short audio single pass

Validated

Tested with an ~8 minute recording on parakeet-tdt-0.6b-v3-int8. 17 chunks, 8.9s total, clean output with no garbled words at boundaries.

[cjpais]

@humblemuzzu would you mind sharing the audio file you have? Or uploading it as part of your commit? Just curious

[humblemuzzu]

that was just my long prompt to claude pretty unhinged but happy to share on X, dmed you