TingMix-Small

CPU-first, natively streaming speech recognition for Taiwanese Mandarin, English and Mandarin-English code-switching. Same streaming API as TingMix-Large; code: github.com/tc3oliver/tingmix.

• Causal log-mel front end -> 2 causal strided convs (25 frames/s) -> unidirectional GRU 768 x 4 -> CTC over 5,025 units (CJK characters + 400 English BPE pieces). Every layer is causal: audio is consumed once, state is kept between chunks, and chunked decoding equals offline decoding.
• Real time on CPU: RTF 0.0604 with 4 threads, 100 ms chunks, p50 5.69 ms / p95 6.85 ms compute per chunk (measured on a server CPU core budget of 4 threads).
• Accuracy is far below TingMix-Large: use Large when a GPU is available.

Use

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install "tingmix[large,server] @ git+https://github.com/tc3oliver/tingmix.git"
tingmix listen --model tc3oliver/TingMix-Small --lm lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz --lm-alpha 0.35 --lm-beta 1.5        # microphone, CPU
tingmix transcribe audio.wav --model tc3oliver/TingMix-Small --lm lm/lmCN_cv_parl_ascend_ntuml.charlm.json.gz --lm-alpha 0.35 --lm-beta 1.5

Partial results are greedy (cheap); with --lm each final result is re-decoded by CTC prefix beam search with the bundled character 5-gram LM (n-gram counts from training transcripts, no audio).

Test results (error rate %, LM decoding of finals, single run)

Test set	Metric	TingMix-Small
Common Voice zh-TW 27.0	CER	30.36
Taiwan Parliament A	CER	26.60
Taiwan Parliament B	CER	26.87
ASCEND, code-switched sentences	MER	43.99
NTUML2021, code-switched sentences (in-domain)	MER	26.47
FLEURS en_us	WER	80.32
LibriSpeech test-clean	WER	39.71
LibriSpeech test-other	WER	60.49
VoxPopuli en	WER	48.18
AMI IHM	WER	64.77

Model chosen on development sets only; scoring as in the Embed2ASR / TingMix-Large benchmark. Training data: same sources as TingMix-Large (not distributed). Trained from scratch (small-s4).