chatterbox-hi-q8-q4-mnn

Bbkblo/chatterbox-hi-q8-q4-mnn

chatterbox-hi-q8-q4-mnn is a text-to-speech model from Bbkblo. Use it when you need text read aloud. It is set up for chatterbox. The card lists the license as apache-2.0.

Source: ResembleAI/Chatterbox-Multilingual-hi (0.5B T3 transformer + S3Gen decoder, fp32 ≈ 2.14 GB).

Downloads · 30 days

Access

Public

Updated Aug 31, 2026

Repo size

1.9 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors1.9 GB · 100%

At a glance

Task: Text-to-Speech
Library: chatterbox
License: apache-2.0
Access: Public
Created: Aug 30, 2026
Updated: Aug 31, 2026
SHA: ba4909f9

Model card

From the Hugging Face model README

Chatterbox-Multilingual-Hindi — Q8 / Q4 quantized (+ MNN toolkit)

Source: ResembleAI/Chatterbox-Multilingual-hi (0.5B T3 transformer + S3Gen decoder, fp32 ≈ 2.14 GB).

Actual model files (this repo)

Path	What	Size
`q8/t3_hi_int8.safetensors`	T3 transformer — INT8 weight quantized (per-output-row quant, int8 w + float scale)	~514 MB
`q4/t3_hi_int4.safetensors`	T3 transformer — INT4 weight quantized (group=128, 2 nibbles packed) + scale	~272 MB
`s3gen_v3.safetensors`	S3Gen V3 speech decoder (kept fp32)	~1008 MB
`grapheme_mtl_merged_expanded_v1.json`	multilingual tokenizer config	~0.1 MB

These are the weight-level quantized checkpoints. Each quantized linear/conv weight tensor is stored as int8 (or packed int4) alongside a matching weight_scale tensor. LayerNorms/embeddings are kept fp32 for stability.

Size comparison

Variant	T3 weight size	Ops
fp32 (base)	~2.0 GB	—
Q8 / INT8	~514 MB	int8 × float scale
Q4 / INT4	~272 MB	group-scale, packed

Quantization scripts

scripts/quantize_weights.py — int8 / int4 (unpacked)
scripts/quant_int4_pack.py — int4 packed (2 nibbles/byte)
scripts/stream_quant.sh / quant_stream.py — streaming per-tensor quantize (low-RAM)

⚠️ Important — about MNN for Chatterbox

Chatterbox ek custom autoregressive TTS hai (T3 token-transformer → S3Gen mel/vocoder). Ye safely one-click ONNX/MNN export nahi hota — T3 ke op graph + autoregressive generation ko MNN mein wirit karne ke liye manual MNN::Express wiring + custom op (RoPE/sinusoidal) implementation chahiye. Isliye:

Is repo mein quantized safetensors (weights) hain — ye weights hain, par MNN .mnn nahi. MNN .mnn ke liye aapko ONNX export → mnnconvert karna padta hai (jisake liye chatterbox package + torch + graph needed, aur is sandbox RAM mein possible nahi).
Woh DhVaani MNN repo jo already mil raha hai (Bbkblo/DhVaani-0.5-MNN) — wo MNN .mnn files real hain aur load/test ho chuke hain (neeche dekho).

For a working MNN TTS right now → use DhVaani

Bbkblo/DhVaani-0.5-MNN (sahi, tested). Chatterbox ka MNN banao to aapko:

pip install chatterbox-tts (≥8GB RAM / GPU pe)
(poore) model mein .t3 + .s3gen → ONNX export
mnnconvert → .mnn

📱 Android Setup Status — READ THIS FIRST

Ye repo MNN nahi hai. Isme sirf quantized .safetensors weights hain. Directly Android TTS app mein nahi lagta.

ANDROID_SETUP.md (repo mein) mein pura honest status + MNN banane ka exact path hai:

ONNX export (torch, ≥8GB RAM/GPU) — model.t3 + model.s3gen

mnnconvert → .mnn

RoPE/sinusoidal + autoregressive custom op wire (MNN MNN::Express)

Abhi Android ke liye ready model → Bbkblo/DhVaani-0.5-MNN (MNN files, tested). Ye repo tab use karo jab aap MNN export + custom op kaam karein.

Reference: https://github.com/resemble-ai/chatterbox · https://github.com/alibaba/MNN

Task: Text-to-Speech
Library: chatterbox
License: apache-2.0
Created: Aug 30, 2026
Updated: Aug 31, 2026