File size: 1,502 Bytes
b345bef
 
af7ca9e
 
b345bef
af7ca9e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
# Streamalign (R16)

Complete R16 speech tokenizer stack. Paired SLM: `Streamalign-SLM-R16`
(reported metrics: SALMon 69.1, StoryCloze 72.1).

## Contents & how the tokenizer uses them

| Path | Role |
|------|------|
| `rvq_teacher/epoch_16.pt` | **The acoustic tokenizer.** Contains the fine-tuned streaming encoder **and** the R16 residual-VQ quantizer (`hubert_state_dict`: 408 encoder keys + `residual_vq`). This is the checkpoint that actually produces the R16 speech units. R16-specific. |
| `alignment_model/` | Char-level streaming Conformer-Transducer (`CKPT+2026-03-15`, `alignment.yaml`). Used as the base model the tokenizer is constructed from; supplies the RNN-T predictor/joiner for char-level alignment. Its encoder is overridden at load by `rvq_teacher`. R-independent. |
| `streaming_asr/` | Word-level streaming Conformer-Transducer (`3419_fastemit`, `chunk_streaming_word_fastemit.yaml`, `tokenizer.ckpt`). Generates the chunk TextGrids (word alignment) and drives the boundary classifier. R-independent. |
| `boundary_classifier/` | Word-boundary detector for streaming chunking; consumes `streaming_asr`. R-independent. |
| `alignment.yaml` | hparams for the extractor / `alignment_model`. |

## Pipeline
audio -> `streaming_asr` (word chunks + TextGrids) + `boundary_classifier`
      -> `alignment_model` scaffold + `rvq_teacher` encoder -> R16 RVQ units.

Only `rvq_teacher/epoch_16.pt` is R16-specific; the ASR / alignment /
boundary components are shared across R8/R16/R32.