improve model card: discoverability tags, benefits, quickstart

README.md

@@ -5,13 +5,21 @@ language:
 library_name: transformers
 pipeline_tag: text-to-speech
 base_model: OpenMOSS-Team/MOSS-TTS-Local-Transformer
+datasets:
+- SPRINGLab/IndicTTS-English
 tags:
 - text-to-speech
 - tts
-- moss-tts
+- speech-synthesis
 - indian-english
-- lora
+- indian-accent
 - voice-agent
+- voice-assistant
+- customer-support
+- conversational
+- moss-tts
+- lora
+- audio
 model-index:
 - name: roxi-tts-pro
   results:
@@ -27,52 +35,46 @@ model-index:
       value: 0.18
 ---
 
-# Roxi-TTS Pro (1.7B), Indian-English voice
+# Roxi-TTS Pro (1.7B): Indian-English text-to-speech
+
+Roxi-TTS Pro is a 1.7B text-to-speech model that speaks in a clear, natural Indian-English
+accent. It is built for customer-support calls and website voice assistants, and it is the
+highest-quality voice in the Roxi line. If you need an Indian-English voice that sounds
+warm, professional, and telephony-ready, start here.
 
-A larger, higher-fidelity member of the Roxi-TTS line. This is a LoRA fine-tune of the
-1.7B MOSS-TTS-Local-Transformer on about 4 hours of a single studio speaker, producing a
-clear Indian-English voice for the VozVox voice-agent platform.
+## Why Roxi-TTS Pro
 
-Compared with the 0.1B roxi-tts-v3.1, this 1.7B version is more intelligible (lower WER)
-and more stable (fewer under-generations), at the cost of generation speed. See the
-Performance section before deploying.
+- Natural Indian-English accent, not a generic English voice with an accent bolted on.
+- Highest intelligibility in the Roxi line: word error rate 0.18 (Whisper-base.en), and
+  strong speaker consistency 0.97 (WavLM-SV).
+- Stable generation with fewer cut-offs than the smaller models, so most lines are usable
+  on the first try.
+- 24 kHz output, single consistent branded voice.
+- Apache-2.0 base models, so it is commercially permissive end to end.
 
-## Model at a glance
+## Quick facts
 
 | Field | Value |
 |---|---|
 | Base model | OpenMOSS-Team/MOSS-TTS-Local-Transformer (1.7B, Apache-2.0) |
-| Audio tokenizer | OpenMOSS-Team/MOSS-Audio-Tokenizer (32 codebooks, Apache-2.0) |
+| Audio tokenizer | OpenMOSS-Team/MOSS-Audio-Tokenizer (Apache-2.0) |
 | Method | LoRA (PEFT), r=32, alpha=64, merged into the base weights |
-| LoRA targets | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
-| Training data | about 4 hours, single IndicTTS-English speaker, 2371 clips |
+| Training data | About 4 hours, single IndicTTS-English speaker, 2371 clips |
 | Output | 24 kHz mono |
 | Speaker similarity | 0.97 (WavLM-SV cosine to held-out target) |
 | Intelligibility WER | 0.18 (Whisper-base.en on generated audio) |
+| Speed | Real-time factor about 2.5 on a 16 GB GPU (best for offline or premium audio) |
 
-## Performance and deployability
-
-Measured on a single 16 GB GPU (bf16, SDPA attention):
-
-- Real-time factor about 2.5, that is roughly 13 seconds of compute per 5 seconds of audio.
-- Peak GPU memory about 13.4 GB during generation.
-
-This model is well suited to offline or pre-rendered speech and to a premium quality tier.
-For live, low-latency turn taking, prefer the 0.1B roxi-tts-v3.1, or optimize this model
-with quantization, torch.compile, a faster GPU, or by caching common phrases.
+## Install
 
-## Requirements
-
-This model uses MOSS-TTS custom code and is built for transformers 4.57.1. Newer versions
-may break generation. Pin the version and install the MOSS-TTS repository so the model class
-is importable.
+Built for transformers 4.57.1. Install the MOSS-TTS repository so the model class is importable.
 
 ```bash
 pip install "transformers==4.57.1" torch torchaudio soundfile librosa peft
 git clone https://github.com/OpenMOSS/MOSS-TTS.git
 ```
 
-## Usage
+## Quick start
 
 ```python
 import sys, torch, soundfile as sf
@@ -105,43 +107,47 @@ audio = processor.decode(out)[0].audio_codes_list[0]
 sf.write("out.wav", audio.float().cpu().numpy(), processor.model_config.sampling_rate)
 ```
 
-Generation is autoregressive and occasionally under-generates. If a clip is short, generate
-two or three times and keep the longest, then trim leading and trailing silence.
-
-Tips: write numbers as words, spell brand names phonetically (for example Voz Vox), avoid raw
-abbreviations, and keep sentences to about twelve words for reliability. Do not raise
-max_new_tokens far above the default, since the codec decode grows quadratically in memory.
+Tips for reliable output: write numbers as words, spell brand names phonetically (for
+example Voz Vox), avoid raw abbreviations, and keep sentences to about twelve words.
+Generation is autoregressive and can occasionally under-generate, so if a clip is short,
+generate two or three times and keep the longest, then trim leading and trailing silence.
+Do not raise max_new_tokens far above the default, since the codec decode grows
+quadratically in memory.
 
-## The Roxi line
+## Which Roxi voice should I use
 
-| Model | Base | Speaker sim | WER | Notes |
+| Model | Base | Best for | Speaker sim | WER |
 |---|---|---|---|---|
-| roxi-tts-v2 | MOSS-TTS-Nano 0.1B | 0.96 | 0.26 | speaker A, about 50 min |
-| roxi-tts-v3 | MOSS-TTS-Nano 0.1B | 0.96 | 0.29 | speaker B, about 70 min |
-| roxi-tts-v3.1 | MOSS-TTS-Nano 0.1B | 0.96 | 0.33 | speaker B, about 4 hours |
-| roxi-tts-pro (this) | MOSS-TTS-Local 1.7B | 0.97 | 0.18 | speaker B, about 4 hours, most intelligible |
-
-## Training
-
-- Data: a single expressive speaker isolated from SPRINGLab/IndicTTS-English via WavLM-SV
-  speaker clustering across dataset shards, about 2371 clips, about 4 hours, studio quality,
-  proper-case transcripts.
-- LoRA r=32 alpha=64 on attention and MLP projections, bf16, learning rate 1e-4 with cosine
-  schedule, gradient accumulation 8, gradient checkpointing, 3 epochs. The epoch-1 checkpoint
-  was selected because later epochs overfit the reading style and raised WER.
-- Evaluation on held-out prompts: speaker similarity with microsoft/wavlm-base-plus-sv,
-  intelligibility with openai/whisper-base.en on the actual generated audio.
+| roxi-tts-pro (this) | MOSS-TTS-Local 1.7B | Highest quality, offline or premium audio | 0.97 | 0.18 |
+| roxi-tts-v3.1 | MOSS-TTS-Nano 0.1B | Real-time, live voice agents | 0.96 | 0.33 |
+
+Use Roxi-TTS Pro when quality matters most and you can pre-render or afford a GPU. Use the
+smaller 0.1B voice when you need real-time, low-latency speech for a live agent.
+
+## Performance and deployability
+
+Measured on a single 16 GB GPU (bf16, SDPA attention): real-time factor about 2.5, that is
+roughly 13 seconds of compute per 5 seconds of audio, with peak GPU memory about 13.4 GB.
+This makes Roxi-TTS Pro well suited to offline or pre-rendered speech and to a premium
+quality tier. For live, low-latency turn taking, prefer the 0.1B roxi-tts-v3.1, or optimize
+this model with quantization, torch.compile, a faster GPU, or by caching common phrases.
+
+## Intended use
+
+Indian-English text to speech for customer-support calls and website voice assistants:
+natural, warm or professional, and telephony aware. Single-speaker branded voice.
 
 ## Limitations
 
 - The training data is read speech, so delivery is somewhat formal rather than fully
   conversational.
-- Not real-time on a single consumer GPU. See the Performance section.
+- Not real-time on a single consumer GPU. See Performance.
 - Stochastic under-generation. Use the retry approach and keep sentences short.
-- Style and emotion control are not reliable. The voice is neutral.
+- Style and emotion control are not reliable. The voice is neutral. For emotion, see
+  roxi-tts-emotion.
 - Requires transformers 4.57.1.
 
-## Attribution and license
+## License and attribution
 
 Released under Apache-2.0. Built on MOSS-TTS-Local-Transformer (Apache-2.0) and its audio
 tokenizer (Apache-2.0). Training data is the IIT-Madras Indic TTS English set accessed via