Lead with the audio: Hear it moves above the fold
Browse files
README.md
CHANGED
|
@@ -18,6 +18,41 @@ A LoRA adapter for [`mispeech/midashenglm-gen`](https://huggingface.co/mispeech/
|
|
| 18 |
trained across all five of the model's capabilities — speech, sound effects,
|
| 19 |
music, ambience and mixed scenes — rather than trading one for another.
|
| 20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
## Why
|
| 22 |
|
| 23 |
Two weak spots in the base model, measured rather than assumed.
|
|
@@ -70,41 +105,6 @@ Caption adherence or realism fell on: sfx clap_text -0.023, ambience clap_text -
|
|
| 70 |
|
| 71 |
**Use this for speech.** For beds, effects and mixed scenes, A/B it against the base model on your own prompts first — the held-out flow loss improved on every capability, and on the scene arms that did not translate.
|
| 72 |
|
| 73 |
-
## Hear it
|
| 74 |
-
|
| 75 |
-
Same prompt, same seed in both arms — `generate()` seeds the global RNG, so the
|
| 76 |
-
solver draws identical noise and every audible difference is the adapter.
|
| 77 |
-
|
| 78 |
-
**Speech, 4-word line — "The river remembers everything." The bucket that went 13.9% → 0.0% WER; the base arm garbles it.**
|
| 79 |
-
|
| 80 |
-
| base | scene-v8 |
|
| 81 |
-
|---|---|
|
| 82 |
-
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/speech-4w__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/speech-4w__V8.wav"></audio> |
|
| 83 |
-
|
| 84 |
-
**Speech, 16-word line — both arms intelligible; the adapter is cleaner and tighter.**
|
| 85 |
-
|
| 86 |
-
| base | scene-v8 |
|
| 87 |
-
|---|---|
|
| 88 |
-
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/speech-16w__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/speech-16w__V8.wav"></audio> |
|
| 89 |
-
|
| 90 |
-
**SFX — footsteps on gravel. Texture and caption match improve; the adapter clip runs shorter, the one real remaining cost.**
|
| 91 |
-
|
| 92 |
-
| base | scene-v8 |
|
| 93 |
-
|---|---|
|
| 94 |
-
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/sfx-gravel__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/sfx-gravel__V8.wav"></audio> |
|
| 95 |
-
|
| 96 |
-
**Music — slow solo piano. The adapter plays longer and closer to the caption at this seed.**
|
| 97 |
-
|
| 98 |
-
| base | scene-v8 |
|
| 99 |
-
|---|---|
|
| 100 |
-
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/music-piano__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/music-piano__V8.wav"></audio> |
|
| 101 |
-
|
| 102 |
-
**Ambience — steady rain on a metal roof. Near-identical by design: this arm held.**
|
| 103 |
-
|
| 104 |
-
| base | scene-v8 |
|
| 105 |
-
|---|---|
|
| 106 |
-
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/ambience-rain__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/ambience-rain__V8.wav"></audio> |
|
| 107 |
-
|
| 108 |
## Training
|
| 109 |
|
| 110 |
| | |
|
|
|
|
| 18 |
trained across all five of the model's capabilities — speech, sound effects,
|
| 19 |
music, ambience and mixed scenes — rather than trading one for another.
|
| 20 |
|
| 21 |
+
## Hear it
|
| 22 |
+
|
| 23 |
+
Same prompt, same seed in both arms — `generate()` seeds the global RNG, so the
|
| 24 |
+
solver draws identical noise and every audible difference is the adapter.
|
| 25 |
+
|
| 26 |
+
**Speech, 4-word line — "The river remembers everything." The bucket that went 13.9% → 0.0% WER; the base arm garbles it.**
|
| 27 |
+
|
| 28 |
+
| base | scene-v8 |
|
| 29 |
+
|---|---|
|
| 30 |
+
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/speech-4w__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/speech-4w__V8.wav"></audio> |
|
| 31 |
+
|
| 32 |
+
**Speech, 16-word line — both arms intelligible; the adapter is cleaner and tighter.**
|
| 33 |
+
|
| 34 |
+
| base | scene-v8 |
|
| 35 |
+
|---|---|
|
| 36 |
+
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/speech-16w__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/speech-16w__V8.wav"></audio> |
|
| 37 |
+
|
| 38 |
+
**SFX — footsteps on gravel. Texture and caption match improve; the adapter clip runs shorter, the one real remaining cost.**
|
| 39 |
+
|
| 40 |
+
| base | scene-v8 |
|
| 41 |
+
|---|---|
|
| 42 |
+
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/sfx-gravel__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/sfx-gravel__V8.wav"></audio> |
|
| 43 |
+
|
| 44 |
+
**Music — slow solo piano. The adapter plays longer and closer to the caption at this seed.**
|
| 45 |
+
|
| 46 |
+
| base | scene-v8 |
|
| 47 |
+
|---|---|
|
| 48 |
+
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/music-piano__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/music-piano__V8.wav"></audio> |
|
| 49 |
+
|
| 50 |
+
**Ambience — steady rain on a metal roof. Near-identical by design: this arm held.**
|
| 51 |
+
|
| 52 |
+
| base | scene-v8 |
|
| 53 |
+
|---|---|
|
| 54 |
+
| <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/ambience-rain__BASE.wav"></audio> | <audio controls src="https://huggingface.co/ModelsLab/midashenglm-gen-wer-lora/resolve/main/samples/ambience-rain__V8.wav"></audio> |
|
| 55 |
+
|
| 56 |
## Why
|
| 57 |
|
| 58 |
Two weak spots in the base model, measured rather than assumed.
|
|
|
|
| 105 |
|
| 106 |
**Use this for speech.** For beds, effects and mixed scenes, A/B it against the base model on your own prompts first — the held-out flow loss improved on every capability, and on the scene arms that did not translate.
|
| 107 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
## Training
|
| 109 |
|
| 110 |
| | |
|