Update README.md

README.md

@@ -50,33 +50,50 @@ or diarization results.
 For server use, configure family `audio_flamingo`, task `asr`, and mode `offline`,
 then submit audio to `/v1/audio/transcriptions` with the instruction in `prompt`.
 
+## Correctness Reference
+
+Official Python BF16 output is the reference. Correctness comparisons and
+performance measurements are separate tests.
+
+| Model | Runtime / weights | Result compared with official Python |
+| --- | --- | --- |
+| Audio Flamingo 3 | Official Python BF16 | Reference |
+| Audio Flamingo 3 | audio.cpp BF16 | Exact transcription and music answer |
+| Audio Flamingo 3 | audio.cpp Q8_0 | Exact transcription; coherent but different music answer |
+| Audio Flamingo 3 | audio.cpp Q4_K | Same transcription words with different wrapper and punctuation; coherent but different music answer |
+| Audio Flamingo Next | Official Python BF16 | Reference |
+| Audio Flamingo Next | audio.cpp BF16 | Exact short transcription; matching words and line boundaries on the three-minute test after timestamps were removed |
+| Audio Flamingo Next | audio.cpp Q8_0 | Same short transcription words with one comma omitted |
+| Audio Flamingo Next | audio.cpp Q4_K | Exact short transcription |
+
+Audio Flamingo Next generates timestamps as text rather than structured
+alignment. Its BF16 three-minute response used the same words and line boundaries
+as Python after timestamps were removed, but the generated timestamps drifted.
+
 ## CUDA Performance and Quantization
 
-The following results use an RTX 5090, the CUDA backend, 8 CPU threads, and the
-same decoded 10-second, 16-kHz mono speech input. Wall time and RTF are from the
-second request in an already-loaded session. Peak VRAM covers the complete
-three-request session and was sampled every 100 ms.
-
-| Model | Weight type | Warm wall time | RTF | Peak VRAM | Transcript compared with BF16 |
-| --- | --- | ---: | ---: | ---: | --- |
-| Audio Flamingo 3 | BF16 | 283.3 ms | 0.0283 | 16,854 MiB | Reference |
-| Audio Flamingo 3 | Q8_0 | 184.9 ms | 0.0185 | 9,956 MiB | Exact |
-| Audio Flamingo 3 | Q4_K | 151.7 ms | 0.0152 | 6,280 MiB | Same words; wrapper and punctuation changed |
-| Audio Flamingo Next | BF16 | 246.3 ms | 0.0246 | 16,890 MiB | Reference |
-| Audio Flamingo Next | Q8_0 | 160.9 ms | 0.0161 | 9,992 MiB | Same words; one comma omitted |
-| Audio Flamingo Next | Q4_K | 129.9 ms | 0.0130 | 6,314 MiB | Exact |
-
-The BF16 Audio Flamingo 3 transcription and music-answer outputs were
-byte-identical to the official Transformers implementation. Audio Flamingo Next
-was also byte-identical on transcription. On a separate three-minute timestamped
-request, its words and line boundaries matched Python after timestamps were
-removed, while generated timestamps drifted.
+The following results use an RTX 5090 and the same decoded 10-second, 16-kHz
+mono speech input. Python uses the official Transformers BF16 model with SDPA.
+audio.cpp uses the CUDA backend and 8 CPU threads. Wall time and RTF are from
+the second request in an already-loaded session. Peak VRAM covers the complete
+process and was sampled every 100 ms.
+
+| Model | Runtime / weights | Warm wall time | RTF | Peak VRAM |
+| --- | --- | ---: | ---: | ---: |
+| Audio Flamingo 3 | Official Python BF16 | 292.4 ms | 0.0292 | 16,766 MiB |
+| Audio Flamingo 3 | audio.cpp BF16 | 283.3 ms | 0.0283 | 16,854 MiB |
+| Audio Flamingo 3 | audio.cpp Q8_0 | 184.9 ms | 0.0185 | 9,956 MiB |
+| Audio Flamingo 3 | audio.cpp Q4_K | 151.7 ms | 0.0152 | 6,280 MiB |
+| Audio Flamingo Next | Official Python BF16 | 249.0 ms | 0.0249 | 16,784 MiB |
+| Audio Flamingo Next | audio.cpp BF16 | 246.3 ms | 0.0246 | 16,890 MiB |
+| Audio Flamingo Next | audio.cpp Q8_0 | 160.9 ms | 0.0161 | 9,992 MiB |
+| Audio Flamingo Next | audio.cpp Q4_K | 129.9 ms | 0.0130 | 6,314 MiB |
 
 Quantized outputs remained coherent on the transcription and music-understanding
-requests. Q8_0 is the balanced default for lower VRAM and faster inference. Q4_K
-offers the lowest tested VRAM, but quantization can change wording, punctuation,
-or factual details in open-ended answers and should not be treated as exact
-parity with BF16.
+requests. Q4_K is the default package and offers the lowest tested VRAM. Q8_0 is
+the higher-precision quantized alternative. Quantization can change wording,
+punctuation, or factual details in open-ended answers and should not be treated
+as exact parity with BF16.
 
 ## Audio Preprocessing