Index-Echo S2TT 2B — GGUF for CrispASR

Conversion of IndexTeam/Index-Echo-S2TT-2B for the index-echo backend in CrispASR. Weights have passed the independent CPU/CUDA checks below. The backend is available on CrispASR main, including C ABI loading, companion downloads and a pinned nightly regression.

The released model transcribes Chinese and translates each subtitle into English, Japanese or Spanish. CrispASR preserves the three-line generation recipe and returns bilingual cues with timestamps and punctuation. English JFK transcription/English translation also matches the released Python model in the CPU fixtures; this is a port-parity result, not a language-wide accuracy claim. The 9B checkpoint is not included.

Files and companions

Both files of the chosen precision must be in the same directory. The tower's index_echo.decoder_file metadata names its decoder; point CrispASR at the tower, not the decoder. The decoder omits the absent MTP branch and contains exactly the 24 text blocks used by the released speech inference path. Vision weights are omitted.

For speech-boundary windows, also put ggml-silero-v6.2.0.bin from ggml-org/whisper-vad beside them, or pass --vad-model. Silero uses its separate MIT license. Without Silero, CrispASR uses bounded 60-second windows. The source recipe keeps context from up to five prior windows and resets history for each input file.

Use

Build CrispASR with the Index-Echo backend, then download a matching pair:

hf download cstr/index-echo-2b-GGUF \
  index-echo-2b-q8_0.gguf index-echo-2b-decoder-q8_0.gguf \
  --local-dir models/index-echo
hf download ggml-org/whisper-vad ggml-silero-v6.2.0.bin \
  --local-dir models/index-echo
./build/bin/crispasr --backend index-echo \
  -m models/index-echo/index-echo-2b-q8_0.gguf \
  -f audio.wav --target-lang en -osrt

--target-lang en|ja|es selects the translation. --prompt accepts glossary entries such as name:translation; --ask overrides the task instruction. Greedy decoding is the default. Positive temperature enables sampling; --seed and --max-new-tokens are forwarded. Custom instructions need the three-line timestamp/transcript/translation format for parsed bilingual cues; unformatted custom output is returned as a whole-window text segment. Word timestamps, beam search and live streaming are not exposed by this recipe.

Validation and quantization

CPU validation run 36905507951 passes F16 and Q8_0 against independent CPU captures (F32 tower/connector, BF16 decoder) from the released Python inference class: English JFK, Chinese and a partial-hop JFK clip. Checks cover mel/convolutions, all 32 audio blocks, the connector, prompt IDs, all 24 decoder blocks, logits and 16 cached teacher-forced steps. The actual Python Session/C ABI loads a filename without a backend hint and reproduces all decoded text and centisecond timestamps exactly.

A separately forced F32 decoder diagnostic now proves higher-precision math: run 36915483750 passes every F16 stage, magnitude, cache, direct-output and full-file gate, including English/Japanese/Spanish and two-window context. Minimum cosine is 0.999992; maximum stage relative L2 is 0.239%. Q8 passes stage/cache/direct checks against that reference too (0.997150 minimum, 3.094% maximum relative L2). The overall diagnostic run is red because Q8 has the documented full-file precision differences; this is not an all-cohort exact-parity claim.

Cosine alone is scale-blind, so magnitude is gated separately. F16 stages require cosine >= .999 (cached logits >= .998) and relative L2 <= 2%; quantized learned stages require cosine >= .99 and relative L2 <= 5%. Frontend stages retain strict gates. Prompt IDs, cached greedy IDs and full direct-window decoded cues are additional mandatory checks. Full-file quality acceptance separately requires the complete text of an independent source precision variant, with exact F16 timestamps and at most 20ms Q8 timestamp drift. No arbitrary text edits are accepted.

Plain Q4_K is rejected: worst cosine 0.781530, relative L2 up to 23.349%, and changed translations/punctuation. Selective Q4 is also rejected: it fails stage/output checks and saves only 2.8% versus Q8. Neither Q4 recipe is a recommended download.

Full-file checks exposed and fixed a Silero waveform-context discrepancy. With that fix, F16 full-file English/Japanese/Spanish and two-window outputs match the separate explicitly F32 decoder source run. The released BF16 decoder differs by 20ms on three Japanese boundaries and chooses a Spanish synonym (creamos versus F32/native F16 pensemos). Q8 also shifts one English boundary by 20ms. Exact mixed-precision parity is therefore claimed only for the three direct-window fixtures, not all full-file multilingual outputs.

A same-host four-core ARM Neoverse-N2 profile shows Q8 warm speedups of 1.33–1.46× over the actual F32/BF16 Python blueprint, with peak RSS 4.03 versus 7.95GiB. A two-Tesla-T4 CUDA run passes stage/cache checks for both cohorts; GPU SRTs match native CPU controls byte for byte. Q8 inference is 4.5–5.3× realtime on those three short clips, excluding load. See CrispASR PERFORMANCE.md and its JSON receipts for methodology and limitations.

Reproducibility and license

Source revision: 5d98a34d9685869b11e9c94d01dee22a8b8e53b5. The tower converter is models/convert-index-echo-to-gguf.py; the standard Qwen3.5 decoder converter is llama.cpp commit 42d958167a748f2c04b1f888e84e7a58f609ddcb, with --outtype f16 --no-mtp. Quantization uses crispasr-quantize; frontend constants and small affines retain floating-point precision. Stage fixtures are in cstr/crispasr-regression-fixtures at 374efe4d7f5c4ff5dce32deffb14e23865fa8493: released mixed-precision captures under index-echo-2b/, separately forced F32 decoder captures under index-echo-2b-f32/. The nightly uses the F32 fixture; the original golden is preserved. Both source full-pipeline captures and effective-dtype audit are included.

The model weights are Apache-2.0; the source license is included as LICENSE. CrispASR runtime code has its own repository license. This conversion is not an official IndexTeam release.