audio-cpp / Index-Echo-S2TT-GGUF

Index-Echo-S2TT GGUF

GGUF packages for native speech-to-text translation with audio.cpp . Chinese speech produces timestamped source transcripts and English, Spanish, or Japanese translations. This port supports S2TT only ; it does not generate translated speech. There is room to improve 9B output parity and quantized quality . Focused PRs with reproducible C++/Python comparisons are welcome.

Note that 2B and 9B use the same C++ translation path, including the prompt, audio encoder, decoder implementation, and greedy sampler. 2B orig is parity safe, while 9B orig has drifts. 2B has 24 layers and tied token/output embeddings; 9B has 32 layers and a separate LM head. That can change how sensitive their outputs are to numerical differences, but I have not proved that it explains the 9B mismatch.

Upstream and Packages

These files were converted from the official 2B checkpoint at 72fd8bdc9bf8251f1c35f6d6289a950eda154395 and 9B checkpoint at 05f86cb7a38684916e193282b6cc02462ace4043 . Each GGUF embeds the audio tower, connector, text decoder, tokenizer, configuration, and audio.cpp model spec. The unused vision weights are not included.

9B Q4_K is not provided. It repeated or omitted speech on short clips and failed a long-form request with incomplete subtitle cues, despite passing the short smoke test. The 2B Q4_K package completed the 20 clips below, but that small set does not establish general reliability. Use Q8_0 or original precision when transcript quality matters. The underlying model can also repeat on some audio even at original precision; inspect subtitles before using them as ground truth.

C++ vs Python: Output Comparison

The test uses 20 distinct Chinese clips from the public FLEURS Chinese test split , with speech windows trimmed using the official Python Silero VAD. This holds the input window fixed across implementations. Source Chinese lines, not translations, were scored against the FLEURS references; WER uses Jieba word segmentation and CER uses normalized characters. This is a small diagnostic slice, not a broad model-quality benchmark. The WER/CER table measures source transcription only; it does not measure translation quality.

Clip zh_016 caused repetition for both Python 9B and C++ 9B variants and is excluded from their scores. With BF16 decoder activation rounding, C++ 2B original follows Python's five-cue output on zh_000 , including its repeated ending. The lower Q8_0/Q4_K WER is mostly due to those variants ending after two cues on that clip; it is not evidence of better transcription quality.

We also compared the saved subtitle cues directly. The 2B original row uses all 20 clips; the 2B quantized rows exclude zh_000 , and all 9B rows exclude zh_016 . Each pair below has the same cue count and is matched in order. Text agreement ignores case, spaces, and punctuation; timestamp agreement counts start and end points within 0.1 seconds.

These are output-agreement checks, not an independent translation-quality score or a claim of byte-identical parity. With the same VAD-trimmed inputs and greedy decoding, the 2B original package's 52/52 matching source cues and 103/104 close timestamp endpoints provide strong evidence that the main model math and decoding logic agree with Python. Its translations still differ in 3/52 cues, so this does not establish identical intermediate calculations. In particular, 2B Q4_K's similar source WER hides more translation changes.

For 9B orig and the zh_004 test case, Python keeps the second phrase in one cue while C++ splits it into two; the largest timestamp difference reflects that boundary shift, not missing audio.

Ordinary CUDA Performance

• NVIDIA RTX 5090; audio.cpp Debug build and official PyTorch inference. One 6.2-second request warmed the loaded model, followed by one 180.3-second request . The long WAV concatenates the 20 distinct Chinese FLEURS speech clips above; it is a long-form workload, not a natural continuous recording.
• The table times only that second, long request, excluding model loading and the warmup. RTF is its wall time divided by 180.3 seconds. C++ uses its quiet-energy windowing; official Python uses Silero VAD, so their window boundaries and generated cue counts are not identical.
• Python used its packaged default BF16 loading. C++ used each GGUF's stored weight type; original-precision packages also round decoder activations to BF16, while quantized packages retain their existing execution path. Neither run forced TF32 or used parity-only controls. Both used greedy decoding.
• Peak VRAM was sampled with nvidia-smi every 50 ms across loading, warmup, and the long request, with no other GPU process running. It is device memory used, not just the PyTorch allocator's tensor count.

All completed C++ runs produced subtitle cues through approximately 03:00. These figures describe this input and machine, not a general latency or accuracy guarantee.

Usage

audiocpp_cli --task asr --family index_echo \
  --model /path/to/Index-Echo-S2TT-GGUF/index-echo-s2tt-9b-q8_0.gguf \
  --backend cuda --audio input.wav \
  --request-option target_language=en \
  --text-out subtitles.txt --log

Set target_language to en , es , or ja . For long input, audio.cpp uses windowed inference and carries the previous five windows of transcript and translation as context. Its default quiet-energy windowing is not identical to the official Python Silero VAD, so boundaries and output may differ. See the model guide for conversion and request options.

License

The weights retain the upstream Apache-2.0 license . Conversion does not replace the upstream license terms.