---
language:
- en
license: apache-2.0
tags:
- token-classification
- prosody
- tobi
- speech
- linguistics
- distilbert
base_model: distilbert-base-uncased
datasets:
- LibriTTS
- peoples_speech
- SBCSAE
---

# text2tobi — `libri+peoples+sbc`

ToBI prosodic annotation from text alone. No audio required at inference time.

Given a stream of lowercased words, this model predicts:
- **Intonation unit boundaries** — where a prosodic phrase ends
- **Intonation direction** at each boundary — rising (H%), falling (L%), or level (!H%)
- **Break index strength** at each boundary — intermediate (3) or full (4)

This is the `libri+peoples+sbc` checkpoint: the best-performing configuration from the Text2ToBI experiments, trained on LibriTTS + People's Speech + SBCSAE with boundary loss weight 2.0, no POS injection, punctuation stripped.

---

## Usage

The recommended way to use this model is via the [text2tobi CLI](https://github.com/your-handle/text2tobi):

```bash
pip install torch transformers huggingface_hub
git clone https://github.com/your-handle/text2tobi
cd text2tobi
python -m text2tobi download
python -m text2tobi "the students filed into the lecture hall"
```

Example output (default table format):

```
word        boundary    intonation    break_index
the         -           -             -
students    -           -             -
filed       -           -             -
into        -           -             -
the         -           -             -
lecture     -           -             -
hall        B           L%            4
```

Pass `--raw` for inline annotations or `--ssml` for SSML XML output.

### Loading directly

If you want to load the model without the CLI, include `model.py` from this repo in your working directory:

```python
from model import ProsodyBoundaryModel
from transformers import AutoTokenizer

model = ProsodyBoundaryModel.from_pretrained("your-handle/text2tobi")
tokenizer = AutoTokenizer.from_pretrained("your-handle/text2tobi")
model.eval()
```

The model returns a dict of logits keyed `boundary_logits`, `intonation_logits`, and `break_idx_logits`.

---

## Performance

Evaluated on SBC001–005 (held-out test set, never seen during training). This is the only configuration directly comparable to the GPT-Neo text-only baseline from Roll et al. (2023).

| Model | Boundary F1 | Intonation F1 | Break Index F1 |
|---|---|---|---|
| **text2tobi** `libri+peoples+sbc` BLW=2.0 | **0.8352** | **0.5765** | 0.6018† |
| GPT-Neo 1.2B (Roll et al., 2023) | 0.770 | — | — |
| Random (distribution-matched) | 0.257 | — | — |

†Break index F1 is evaluated on BU Radio News Corpus gold `.brk` annotations (not the SBC test set, which has no break index labels). Treat as experimental.

text2tobi surpasses the GPT-Neo baseline by 6.5 points while being approximately 18× smaller (~66M vs ~1.2B parameters), and without access to punctuation or capitalization — input is lowercased words only.

---

## Training data

| Corpus | Annotation | Role |
|---|---|---|
| LibriTTS | Silver (PSST + Wav2ToBI consensus) | Boundary + intonation |
| People's Speech | Silver (PSST + Wav2ToBI consensus) | Boundary + intonation |
| SBCSAE | Gold (Du Bois transcripts) | Boundary + intonation |
| BU Radio News | Gold (`.brk` files) | Break index evaluation only |

Silver-standard boundary and intonation labels were generated by cross-validating PSST (`NathanRoll/psst-medium-en`) against Wav2ToBI (`ReginaZ/Wav2ToBI-PB-Fuzzy`). Positions where the two systems disagreed were masked from training. 87.3% of utterance-final words received Wav2ToBI corroboration within ±1 word.

SBCSAE data is included under explicit written permission from corpus director John W. Du Bois (June 2026) for unrestricted public distribution of derived model weights.

---

## Known limitations

- **Intonation labels apply to boundary words only.** Non-boundary intonation is not modeled.
- **Register coverage** is read speech (LibriTTS, People's Speech) and conversational speech (SBCSAE). Generalization to telephony, noisy environments, or non-native speakers has not been tested.
- **Chunking fallback**: for unpunctuated input, the inference pipeline splits at a 100-token word boundary when no sentence boundary is detected. This is not linguistically motivated and may affect predictions near split points.

---

## License

Apache 2.0. See [LICENSE](LICENSE).