Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

test: numerical regression harness (frozen golden vs ggml/VAD/CIF/CTC output) (#3003)

zhifu gao3610b759ce9e
diff --git a/runtime/llama.cpp/tests/README.md b/runtime/llama.cpp/tests/README.md
new file mode 100644
index 000000000..0d7d9da34
--- /dev/null
+++ b/runtime/llama.cpp/tests/README.md
@@ -0,0 +1,18 @@
+# Regression tests
+
+Run each runtime tool on a fixed clip (`sample.wav`, ~6 s) and diff the output against
+the frozen golden in `golden/`. Catches regressions in the ggml graphs, the FSMN-VAD
+state machine, the CIF predictor and CTC decode.
+
+## Run
+```bash
+# from runtime/llama.cpp/, after building (cmake --build build):
+./tests/run_regression.sh # VAD (tiny model auto-fetched) + any tool whose GGUF is already local
+RUN_FULL=1 ./tests/run_regression.sh # also download the ASR GGUFs from Hugging Face and test every tool
+```
+`BIN_DIR` / `MODELS_DIR` override where binaries and GGUFs are found. Exit code is
+non-zero if any test fails; tools with no binary or no model are skipped.
+
+## Golden
+Captured on Linux x86-64 (the reference platform) with the f16 GGUFs published at
+`FunAudioLLM/*-GGUF`. Update a golden file only with a deliberate, reviewed output change.
diff --git a/runtime/llama.cpp/tests/golden/nano.txt b/runtime/llama.cpp/tests/golden/nano.txt
new file mode 100644
index 000000000..bcda42723
--- /dev/null
+++ b/runtime/llama.cpp/tests/golden/nano.txt
@@ -0,0 +1 @@
+我想问我在滨海新区有房
diff --git a/runtime/llama.cpp/tests/golden/paraformer.txt b/runtime/llama.cpp/tests/golden/paraformer.txt
new file mode 100644
index 000000000..bcda42723
--- /dev/null
+++ b/runtime/llama.cpp/tests/golden/paraformer.txt
@@ -0,0 +1 @@
+我想问我在滨海新区有房
diff --git a/runtime/llama.cpp/tests/golden/sensevoice.txt b/runtime/llama.cpp/tests/golden/sensevoice.txt
new file mode 100644
index 000000000..bcda42723
--- /dev/null
+++ b/runtime/llama.cpp/tests/golden/sensevoice.txt
@@ -0,0 +1 @@
+我想问我在滨海新区有房
diff --git a/runtime/llama.cpp/tests/golden/vad.txt b/runtime/llama.cpp/tests/golden/vad.txt
new file mode 100644
index 000000000..d9f88ff55
--- /dev/null
+++ b/runtime/llama.cpp/tests/golden/vad.txt
@@ -0,0 +1 @@
+770 5980
diff --git a/runtime/llama.cpp/tests/run_regression.sh b/runtime/llama.cpp/tests/run_regression.sh
new file mode 100755
index 000000000..50eab5d5e
--- /dev/null
+++ b/runtime/llama.cpp/tests/run_regression.sh
@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# Numerical regression test for the FunASR llama.cpp runtime.
+# Runs each available tool on a fixed clip (sample.wav) and diffs the output against
+# the frozen golden in golden/. Catches regressions in the ggml graphs, the FSMN-VAD
+# state machine, CIF predictor and CTC decode. Golden captured on Linux x86-64 (the
+# reference platform) with the f16 GGUFs published on Hugging Face.
+#
+# ./run_regression.sh # test tools whose binary+model are present (VAD model auto-fetched)
+# RUN_FULL=1 ./run_regression.sh # also download the ASR GGUFs and test every tool
+# BIN_DIR=/path/to/bin MODELS_DIR=/path/to/gguf ./run_regression.sh
+set -u
+DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+RT=$(cd "$DIR/.." && pwd)
+BIN="${BIN_DIR:-$RT/build/bin}"
+MODELS="${MODELS_DIR:-$DIR/models}"
+SAMPLE="$DIR/sample.wav"
+RUN_FULL="${RUN_FULL:-0}"
+mkdir -p "$MODELS"
+pass=0; fail=0; skip=0
+
+bin(){ if [ -x "$BIN/$1" ]; then echo "$BIN/$1"; elif command -v "$1" >/dev/null 2>&1; then echo "$1"; fi; }
+# ensure the listed model files exist in $MODELS (download via the repo script if allowed)
+ensure_models(){ local key="$1"; shift
+ local missing=0; for f in "$@"; do [ -f "$MODELS/$f" ] || missing=1; done
+ [ "$missing" = 0 ] && return 0
+ { [ "$key" = fsmn-vad ] || [ "$RUN_FULL" = 1 ]; } || return 1 # only auto-fetch the tiny VAD unless RUN_FULL
+ "$RT/download-funasr-model.sh" "$key" "$MODELS" >/dev/null 2>&1 || return 1
+ for f in "$@"; do [ -f "$MODELS/$f" ] || return 1; done; return 0
+}
+check(){ local name="$1" golden="$2" got="$3"
+ if [ "$got" = "$(cat "$golden")" ]; then echo " PASS $name"; pass=$((pass+1))
+ else echo " FAIL $name"; echo " expected: $(head -c 80 "$golden")"; echo " got: $(printf %s "$got" | head -c 80)"; fail=$((fail+1)); fi
+}
+skipper(){ echo " SKIP $1 ($2)"; skip=$((skip+1)); }
+
+run_tool(){ # name binary golden key models... -- run...
+ local name="$1" b key gold; b=$(bin "$2"); gold="$DIR/golden/$3"; key="$4"; shift 4
+ local models=(); while [ "$1" != "--" ]; do models+=("$1"); shift; done; shift
+ [ -n "$b" ] || { skipper "$name" "no binary"; return; }
+ [ -f "$gold" ] || { skipper "$name" "no golden"; return; }
+ ensure_models "$key" "${models[@]}" || { skipper "$name" "model missing (set RUN_FULL=1)"; return; }
+ check "$name" "$gold" "$("$@" 2>/dev/null)"
+}
+
+echo "== FunASR llama.cpp regression (sample.wav) =="
+B=$(bin llama-funasr-vad) && run_tool vad llama-funasr-vad vad.txt fsmn-vad fsmn-vad.gguf -- "$B" -m "$MODELS/fsmn-vad.gguf" -a "$SAMPLE"
+B=$(bin llama-funasr-sensevoice) && run_tool sensevoice llama-funasr-sensevoice sensevoice.txt sensevoice sensevoice-small-f16.gguf -- "$B" -m "$MODELS/sensevoice-small-f16.gguf" -a "$SAMPLE"
+B=$(bin llama-funasr-paraformer) && run_tool paraformer llama-funasr-paraformer paraformer.txt paraformer paraformer-f16.gguf -- "$B" -m "$MODELS/paraformer-f16.gguf" -a "$SAMPLE"
+B=$(bin llama-funasr-cli) && run_tool nano llama-funasr-cli nano.txt nano funasr-encoder-f16.gguf qwen3-0.6b-q8_0.gguf -- "$B" --enc "$MODELS/funasr-encoder-f16.gguf" -m "$MODELS/qwen3-0.6b-q8_0.gguf" -a "$SAMPLE"
+
+echo "== $pass passed, $fail failed, $skip skipped =="
+[ "$fail" = 0 ]
diff --git a/runtime/llama.cpp/tests/sample.wav b/runtime/llama.cpp/tests/sample.wav
new file mode 100644
index 000000000..f5db87cf6
Binary files /dev/null and b/runtime/llama.cpp/tests/sample.wav differ