Eval harness v1: regenerate eval set with stock + SOAR heads, Whisper WER + resemblyzer SIM-O + drift, push to eval repo

eval_2x2.py

@@ -0,0 +1,167 @@
+"""eval_2x2.py — Eval harness for the 2×2 ablation: stock vs SOAR head on identical scripts.
+
+Generates the same 9 two-speaker dialogue scripts (short/medium/long × 3) with BOTH the
+stock head and the SOAR-trained head (ACloudCenter/vibevoice-soar-head-1.5B), then:
+- WER per arm via openai/whisper-large-v3 vs the known ground-truth scripts
+- SIM-O per arm: resemblyzer speaker-embedding cosine (reference voice vs generated)
+- Drift: cosine(first 30s, last 30s) within each long sample
+- RTF per sample (revalidates the baseline measurement with the SOAR head loaded)
+
+Env: ARMS (comma list, default "stock,soar")
+Pushes wavs + results.json to ACloudCenter/vibevoice-eval-2x2 under {arm}/.
+Prints EVAL DONE at the end.
+"""
+import io, json, os, sys, time
+sys.path.insert(0, "/VibeVoice-classic")
+
+import numpy as np
+import soundfile as sf
+import torch
+from huggingface_hub import hf_hub_download, HfApi
+
+from vibevoice.modular.modeling_vibevoice import VibeVoiceForConditionalGeneration
+from vibevoice.processor.vibevoice_processor import VibeVoiceProcessor
+
+MODEL = "microsoft/VibeVoice-1.5B"
+SOAR_REPO = "ACloudCenter/vibevoice-soar-head-1.5B"
+OUT_REPO = "ACloudCenter/vibevoice-eval-2x2"
+SR = 24000
+device = "cuda"
+ARMS = os.environ.get("ARMS", "stock,soar").split(",")
+REPO_TYPE = "model"
+
+# ---------------- fixed eval scripts (2 speakers: Speaker 1 / Speaker 2) ----------------
+SHORTS = [
+    "Speaker 1: Did you see the meteor shower last night?\nSpeaker 2: I missed it! Was it worth staying up for?\nSpeaker 1: Absolutely. The sky lit up green for a full minute.\nSpeaker 2: Okay, now I'm actually upset I slept through it.",
+    "Speaker 1: The bread machine finally arrived.\nSpeaker 2: So you're a baker now?\nSpeaker 1: I'm a disaster with a bread machine. The first loaf came out like a brick.\nSpeaker 2: Send me a photo. Immediately.",
+    "Speaker 1: My flight lands at six, so don't wait for dinner.\nSpeaker 2: I'll keep something warm anyway.\nSpeaker 1: You really don't have to.\nSpeaker 2: I know. That's why I'm doing it.",
+]
+MEDIUMS = [
+    "Speaker 1: I've been reading about how old lighthouse keepers kept journals.\nSpeaker 2: What did they write about?\nSpeaker 1: Mostly weather, ships passing, and how lonely it was.\nSpeaker 2: That sounds peaceful and terrible at the same time.\nSpeaker 1: Exactly. One keeper wrote that the light was his only conversation for months.\nSpeaker 2: Honestly, that might fix my screen time problem.\nSpeaker 1: Or break you completely. There was no off switch on the ocean.",
+    "Speaker 1: The city is redoing the bike lane on Fifth Street.\nSpeaker 2: Finally. That stretch is dangerous.\nSpeaker 1: They're adding a physical barrier between cars and cyclists.\nSpeaker 2: About time. I've almost been clipped twice this month.\nSpeaker 1: Construction starts Monday, so expect detours.\nSpeaker 2: Worth it if it means I stop white-knuckling my handlebars.",
+    "Speaker 1: My grandmother used to make bread every Sunday morning.\nSpeaker 2: Sourdough?\nSpeaker 1: A dense rye. She said the recipe came from her grandmother.\nSpeaker 2: Four generations of the same loaf.\nSpeaker 1: I still have her mixing bowl. It's older than the house.\nSpeaker 2: That bowl has earned a museum display.",
+]
+LONGS = [
+    "Speaker 1: Let's talk about sleep. Everyone says eight hours, but nobody agrees on anything else.\nSpeaker 2: Right, and half the advice contradicts the other half.\nSpeaker 1: Some people swear by cold rooms. Others say the temperature barely matters.\nSpeaker 2: The one thing that's consistent is the schedule. Same bedtime, same wake time.\nSpeaker 1: Even on weekends, which is where everyone cheats.\nSpeaker 2: That's exactly where I cheat.\nSpeaker 1: The researchers say your body doesn't know it's Saturday.\nSpeaker 2: My body absolutely knows it's Saturday.\nSpeaker 1: Apparently not, biologically speaking.\nSpeaker 2: Fine, I'll set an alarm for Sunday too. This conversation is costing me my weekend.\nSpeaker 1: Sleep hygiene comes with sacrifices.\nSpeaker 2: Don't lecture me at nine on a Sunday morning.",
+    "Speaker 1: The bakery on the corner closed last week.\nSpeaker 2: No! That place had the best almond croissants in the city.\nSpeaker 1: Rent tripled. The owner said she couldn't make the numbers work anymore.\nSpeaker 2: Where does a neighborhood bakery even go from there?\nSpeaker 1: She's moving production to a shared kitchen across the river and doing delivery only.\nSpeaker 2: So no more counter, no more coffee, just a website?\nSpeaker 1: Just a website. She seemed okay about it, which surprised me.\nSpeaker 2: Thirty years behind that counter and she seemed okay?\nSpeaker 1: Maybe she's more practical than sentimental.\nSpeaker 2: I'm going to miss walking past that smell every morning.\nSpeaker 1: Delivery still smells. It just doesn't linger.",
+    "Speaker 1: I finally fixed the shed door.\nSpeaker 2: The one that's been hanging crooked since spring?\nSpeaker 1: Since spring. The hinges were stripped, not the frame, which saved me a full rebuild.\nSpeaker 2: How long did that take?\nSpeaker 1: Four hours, two of which were finding screws that actually fit.\nSpeaker 2: The hardware store run is always the real project.\nSpeaker 1: Two trips. One for the wrong screws, one for the right ones.\nSpeaker 2: Classic. Did you paint it?\nSpeaker 1: Not yet. The wood needs to dry out first, apparently.\nSpeaker 2: So the project continues.\nSpeaker 1: The project continues. But the door closes now, which is more than March could say.",
+]
+
+def load_model_with_head(load_soar):
+    model = VibeVoiceForConditionalGeneration.from_pretrained(
+        MODEL, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True,
+        attn_implementation="sdpa").to(device).eval()
+    if load_soar:
+        sd_path = hf_hub_download(SOAR_REPO, "soar_head_state.pt")
+        sd = torch.load(sd_path, map_location="cpu", weights_only=True)
+        missing, unexpected = model.model.prediction_head.load_state_dict(sd, strict=True), None
+        print(f"loaded SOAR head: {len(sd)} tensors", flush=True)
+    return model
+
+def main():
+    api = HfApi()
+    scripts = ([("short", i, t) for i, t in enumerate(SHORTS)]
+               + [("medium", i, t) for i, t in enumerate(MEDIUMS)]
+               + [("long", i, t) for i, t in enumerate(LONGS)])
+    refs = {}
+    for spk, name in ((1, "en-Carter_man.wav"), (2, "en-Maya_woman.wav")):
+        wav, sr = sf.read(f"/VibeVoice-classic/demo/voices/{name}")
+        assert sr == SR
+        refs[spk] = wav.astype(np.float32)
+
+    proc = VibeVoiceProcessor.from_pretrained(MODEL)
+    asr = torch.device(device)
+    from transformers import pipeline
+    transcriber = pipeline("automatic-speech-recognition", model="openai/whisper-large-v3",
+                           device=asr, torch_dtype=torch.float16)
+    from resemblyzer import VoiceEncoder, preprocess_wav
+    enc = VoiceEncoder(device)
+
+    def gt_text(script):
+        return " ".join(line.split(": ", 1)[1].replace("\n", " ") for line in script.strip().split("\n"))
+
+    def sim_to_ref(wav, ref_key):
+        emb = enc.embed_utterance(preprocess_wav(wav, source_sr=SR))
+        ref = enc.embed_utterance(preprocess_wav(refs[ref_key_for(ref_key_for_i(ref_key=ref_key_for_i:=None))])) if False else None
+        return emb
+
+    results = []
+    os.makedirs("/out", exist_ok=True)
+    for arm in ARMS:
+        arm = arm.strip()
+        print(f"=== ARM {arm} ===", flush=True)
+        model = load_model_with_head(load_soar=(arm == "soar"))
+        model.model.noise_scheduler = model.model.noise_scheduler  # default ddpm
+        model.set_ddpm_inference_steps(num_steps=10)
+        for tier, idx, script in scripts:
+            t0 = time.time()
+            with torch.no_grad():
+                out = model.generate(
+                    text=[script],
+                    voice_samples=[[refs[1], refs[2]]],
+                    cfg_scale=1.3,
+                    generation_config={"do_sample": False},
+                )
+            audio = out.speech_outputs[0]
+            if torch.is_tensor(audio := audio or None) or True:
+                pass
+            gen_s = time.time() - t0
+            wav = audio if isinstance(audio, np.ndarray) else audio.float().cpu().numpy()
+            audio_s = len(wav) / SR
+            rtf = gen_s / max(audio_s, 1e-6)
+            fn = f"{arm}_{tier}_{idx}.wav"
+            sf.write(f"/out/{fn}", wav, SR)
+            hyp = transcriber(f"/out/{fn}", generate_kwargs={"language": "english"})["text"].strip()
+            gt = " ".join(l.split(": ", 1)[1].strip() for l in script.strip().splitlines())
+            gt_words, hyp_words = gt.lower().split(), hyp.lower().split()
+            import jiwer
+            wer = jiwer.wer(gt, hyp)
+            emb_gen = enc.embed_utterance(preprocess_wav(f"/out/{fn}"))
+            ref_key = tier
+            ref_idx = {"short": 0, "medium": 0, "long": 0}[tier]
+            ref_name = {"Speaker 1": "en-Carter_man.wav", "Speaker 2": "en-Maya_woman.wav"}
+            sims = []
+            for spk_line in script.strip().splitlines():
+                spk = spk_line.split(":")[0] if (spk_line := spk_line) else None
+            # simple: SIM vs combined ref (concat refs)
+            ref_wav = np.concatenate([refs[1], np.zeros(SR // 4, dtype=np.float32), refs[2]])
+            emb_ref = enc.embed_utterance(preprocess_wav(ref_wav.astype(np.float32), SR))
+            sim = float(np.dot(emb_gen, ref_emb) / (np.linalg.norm(emb_gen) * np.linalg.norm(ref_emb)))
+            drift = None
+            if tier == "long" and audio_s > 60:
+                first = wav[: 30 * SR].astype(np.float32); last = wav[-30 * SR:].astype(np.float32)
+                e1 = enc.embed_utterance(preprocess_wav(first, SR))
+                e2 = enc.embed_utterance(preprocess_wav(last, SR))
+                drift = float(np.dot(e1, e2) / (np.linalg.norm(e1) * np.linalg.norm(e2)))
+            results.append({"arm": arm, "tier": tier, "idx": idx, "audio_s": audio_s,
+                            "gen_s": gen_s, "rtf": rtf, "wer": float(wer), "sim_ref": sim,
+                            "drift": drift, "hyp": hyp})
+            print(f"{arm} {tier}{idx}: rtf {rtf:.3f} wer {wer:.3f} sim {sim:.3f} "
+                  f"drift {drift if drift is None else round(drift, 3)}", flush=True)
+            trackio.log({"rtf": rtf, "wer": float(wer), "sim": sim}, step=len(results))
+            api.upload_file(path_or_fileobj=f"/out/{fn}", path_in_repo=f"{arm}/{fn}",
+                            repo_id=OUT_REPO, repo_type="model")
+            torch.cuda.empty_cache()
+        del model
+        torch.cuda.empty_cache()
+
+    with open("/out/results.json", "w") as f:
+        json.dump(results, f, indent=2)
+    import statistics as st
+    summary = {}
+    for arm in ARMS:
+        rows = [r for r in results if r["arm"] == arm]
+        summary[arm] = {"mean_wer": st.mean(r["wer"] for r in rows),
+                        "mean_sim": st.mean(r["sim_ref"] for r in rows),
+                        "mean_rtf": st.mean(r["rtf"] for r in rows),
+                        "long_drift": st.mean(r["drift"] for r in rows if r["drift"] is not None)}
+    with open("/out/summary.json", "w") as f:
+        json.dump(summary, f, indent=2)
+    print("SUMMARY " + json.dumps(summary), flush=True)
+    api.upload_file(path_or_fileobj="/out/results.json", path_in_repo="results.json",
+                    repo_id=OUT_REPO, repo_type="model")
+    api.upload_file(path_or_fileobj="/out/summary.json", path_in_repo="summary.json",
+                    repo_id=OUT_REPO, repo_type="model")
+    print("EVAL DONE", flush=True)
+
+main()