Indic-Mio-LiteRT

LiteRT (.tflite) graph bundle of SPRINGLab/Indic-Mio (Qwen3-0.6B speech-token LM, Apache-2.0) with its Aratako/MioCodec-25Hz-24kHz codec (MIT), for CPU inference on Windows/Linux x86_64 through speech-core. Emotion is controlled with end-of-utterance suffix tags; voice cloning conditions on a 128-dim global speaker embedding extracted from reference audio.

LiteRT cannot express the autoregressive loop, so the bundle ships static graphs and the host owns the loop.

Graphs

File	Inputs	Outputs
`indicmio-text-prefill.tflite` (fp16)	`ids[1,64]` i64 (right-pad with 151643), `last_index` scalar i64	`logits[1,164480]` f32, `k/v[28,1,8,512,128]` f32
`indicmio-token-step.tflite` (fp16)	`id[1,1]`, `pos[1,1]`, `write_index` scalar, `k/v in`	`logits[1,164480]`, `k/v out`
`indicmio-audio-decoder.tflite` (fp32)	`codes[1,384]` i64 (zero-pad), `global[1,128]` f32, `valid_tokens` scalar i64	STFT `real[1,961,768]`, `imag[1,961,768]` f32
`indicmio-ref-encoder.tflite` (fp32)	`audio24k[1,240000]` f32 (10 s bucket)	`global[1,128]` f32

config.json carries the manifest (token offsets, stop ids, prompt template, bucket sizes); fidelity.json the parity results; the HF tokenizer files are included for host-side tokenization.

Host contract

Prompt: <|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant\n, right-padded to 64 tokens with 151643; pass the real last position as last_index. Emotion tags go at the end of the sentence: <happy> <sad> <angry> <disgust> <fear> <surprise>.
AR loop: sample from logits (upstream reference: temperature 0.9, top-p 0.9; suppress EOS until the first speech token), feed the sampled id back through token-step with pos = write_index = current length. Stop on 151645 or 151643. Speech codes are id - 151669 (12800 codes, 25 Hz, 960 samples/token at 24 kHz).
Decode: zero-pad the codes to 384, set valid_tokens to the real count (the graph masks attention and normalization statistics so padding cannot contaminate the result), then reconstruct audio from the spectrum frames: complex spec = real + i·imag, inverse rFFT per frame (n_fft 1920), Hann window, overlap-add at hop 480, divide by the (constant, precomputable) window envelope, trim (win-hop)/2 at both ends and keep valid_tokens*960 samples, guarding the final ~8 hops of the tail.
Cloning: crop/pad the reference to 10 s at 24 kHz and run the ref encoder (prefer cropping — zero-padding slightly dilutes the embedding); pass the result as global. Without a reference, pass zeros.

Parity (vs the upstream PyTorch stack)

Check	Result
Prefill logits corr	~1.000000
Token-step logits corr (non-zero write index)	~1.000000
Decoder PCM corr vs exact-length upstream decode	0.998
Reference-encoder cosine	0.999
Hindi ASR roundtrip (keyword recovery, 4 cases)	pass; emotion tags styled, never spoken

Licensing

The language model derives from SPRINGLab/Indic-Mio (Apache-2.0); the codec weights derive from Aratako/MioCodec-25Hz-24kHz (MIT) and its WavLM-base+ speaker path. This bundle preserves upstream weights (fp16/fp32) with no fine-tuning.

soniqo/Indic-Mio-LiteRT