Shubhangi7/pocket-tts-checkpoints

Pocket TTS: English + Hindi + Hinglish checkpoints

One Pocket TTS model that speaks English, Hindi (Devanagari) and Hinglish (Devanagari with English words in Latin script), finetuned from Kyutai's english_2026-04_24l 24-layer model.

Folders

Folder	Training step	Contents
epoch-5 … epoch-15	epoch × 10,569	model.safetensors (inference weights: EMA FlowLM + Mimi, float32, ~1.3 GB) and config.yaml
epoch-10-distilled	epoch 10, distilled 200,000 steps	6-layer model distilled from epoch-10 (guidance baked in): model.safetensors (~0.44 GB, EMA) and config.yaml. ~3.5× faster than the 24-layer models on CPU
final	250,000 (~23.7 epochs, end of training)	the full training checkpoint as-is: checkpoint_00250000.pt (weights + EMA), optim_00250000.pt (optimizer state, to resume training), model.safetensors, args.yaml (training config), config.yaml
tokenizer.json	–	SentencePiece unigram tokenizer (4,000 pieces) shared by all folders

Usage

pip install pocket-tts
pocket-tts generate --config hf://Shubhangi7/pocket-tts-checkpoints/epoch-10/config.yaml \
    --voice your_voice.wav --text "आपका order ship हो गया है, आप इसे app में track कर सकते हैं।"

from pocket_tts import TTSModel

model = TTSModel.load_model(config="hf://Shubhangi7/pocket-tts-checkpoints/epoch-10/config.yaml")
voice = model.get_state_for_audio_prompt("your_voice.wav")  # 5–10 s of clean speech
audio = model.generate_audio(voice, "Rahul loves cricket और वो हर Sunday match देखता है।")

Streaming works as in Pocket TTS (model.generate_audio_stream(...)). Released pocket-tts splits long text only at ., ! and ?, and appends a . after a final "।"; treating "।" and "॥" as sentence ends needs the small text_chunking.py change from the training repo.

Training

• Data: ~1,429 h: English ~698 h (16 kHz, conversational Indian English), Hindi ~697 h (16 kHz), Hinglish ~36 h (44.1 kHz). Transcripts used verbatim; word timings from the MMS forced aligner (MahmoudAshraf/mms-300m-1130-forced-aligner) with uroman romanisation.
• Recipe: training/configs/finetune_language.yaml from the Pocket TTS repo: start from the English 24-layer teacher, fresh text embedding for the new tokenizer, LSD flow objective, lr 2e-4 constant, batch 64, 250k steps, EMA 0.999. One AMD MI300X, ~18 h.
• Model size: 316M FlowLM parameters (24 layers, d_model 1024) + 20M Mimi codec.
• epoch-10-distilled: the repo's depth_distill.yaml recipe with epoch-10 as the frozen teacher: a 6-layer student (89M FlowLM parameters) seeded from teacher layers 0, 1, 2, 21, 22, 23, regressing the teacher's CFG-2.0 backbone output on the same data; lr 4e-4 cosine, batch 64, 200k steps, EMA 0.9999.

Notes

• Apart from epoch-10-distilled, these are 24-layer models, heavier than the released 6-layer Pocket TTS: on a 24-core CPU, ~4 concurrent real-time streams without buffering. The distilled model runs ~4.7× real time on one CPU thread (24-layer: ~1.3×).
• Much of the English and Hindi audio is 16 kHz, so output tends to follow the bandwidth of the voice prompt; a clean, full-band prompt gives the best quality.
• The Mimi codec weights come from Kyutai's english_2026-04_24l release; their terms apply.