multimodalart HF Staff commited on
Commit
8a287cc
·
verified ·
1 Parent(s): 76460c4

Upload folder using huggingface_hub

Browse files
Files changed (3) hide show
  1. README.md +1 -0
  2. app.py +9 -18
  3. requirements.txt +0 -1
README.md CHANGED
@@ -27,6 +27,7 @@ Native 7.5 Hz duplex streaming doesn't fit a request/response Space, so this dem
27
  ## Tips
28
 
29
  - The reply is the model's full timeline, not just its speaking turns — it includes backchannels (short "mm-hm" acknowledgements) while you talk.
 
30
  - Best voice-clone results come from a clean, single-speaker reference with no music or noise.
31
  - `speak_first` in Advanced settings makes Haru open the conversation instead of waiting.
32
 
 
27
  ## Tips
28
 
29
  - The reply is the model's full timeline, not just its speaking turns — it includes backchannels (short "mm-hm" acknowledgements) while you talk.
30
+ - After your clip the demo appends a 5-second pause, so Haru gets the floor and its spoken answer lands inside the returned timeline.
31
  - Best voice-clone results come from a clean, single-speaker reference with no music or noise.
32
  - `speak_first` in Advanced settings makes Haru open the conversation instead of waiting.
33
 
app.py CHANGED
@@ -12,23 +12,6 @@ os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
12
 
13
  import spaces # ZeroGPU: must come before torch / transformers
14
 
15
- # Browsers record Opus in WebM containers, which soundfile cannot decode; the
16
- # haru package falls back to an `ffmpeg` binary. The Gradio image does not ship
17
- # one on PATH, so expose the static binary bundled with imageio-ffmpeg.
18
- try:
19
- import imageio_ffmpeg
20
-
21
- _source = imageio_ffmpeg.get_ffmpeg_exe()
22
- _bindir = "/tmp/ffmpeg-bin"
23
- os.makedirs(_bindir, exist_ok=True)
24
- _target = os.path.join(_bindir, "ffmpeg")
25
- if not os.path.exists(_target):
26
- shutil.copy2(_source, _target)
27
- os.chmod(_target, 0o755)
28
- os.environ["PATH"] = _bindir + os.pathsep + os.environ.get("PATH", "")
29
- except Exception as _ffmpeg_error: # not fatal: WAV/FLAC/Ogg decode natively
30
- print(f"[boot] ffmpeg shim not set up: {_ffmpeg_error!r}", flush=True)
31
-
32
  import json
33
  import shutil
34
  import time
@@ -43,6 +26,7 @@ from haru.prompt import card_templates, render_card
43
 
44
  MODEL_ID = "higurashi-ai/Haru-4B"
45
  MAX_USER_SECONDS = 30 # duplex turns are short; bounds the ZeroGPU duration
 
46
 
47
  # Character cards: built-in per-language presets shipped inside the haru package.
48
  _TEMPLATES = card_templates()
@@ -62,7 +46,7 @@ model = load_model(MODEL_ID, device="cpu").to("cuda")
62
  print("[boot] Haru-4B ready on cuda", flush=True)
63
 
64
 
65
- @spaces.GPU(duration=90)
66
  def converse(
67
  user_audio,
68
  reference_audio,
@@ -102,6 +86,11 @@ def converse(
102
  limit = MAX_USER_SECONDS * SAMPLE_RATE
103
  trimmed = len(user) > limit
104
  user = user[:limit]
 
 
 
 
 
105
  count = max(1, -(-len(user) // frame_samples))
106
  user = np.pad(user, (0, count * frame_samples - len(user)))
107
 
@@ -151,6 +140,8 @@ def converse(
151
  text = "".join(frame.text for frame in frames).strip()
152
  if trimmed:
153
  text = f"(input trimmed to {MAX_USER_SECONDS} s)\n{text}"
 
 
154
  return (SAMPLE_RATE, reply), text
155
 
156
 
 
12
 
13
  import spaces # ZeroGPU: must come before torch / transformers
14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15
  import json
16
  import shutil
17
  import time
 
26
 
27
  MODEL_ID = "higurashi-ai/Haru-4B"
28
  MAX_USER_SECONDS = 30 # duplex turns are short; bounds the ZeroGPU duration
29
+ PAUSE_SECONDS = 5 # appended silence so Haru's answer lands inside the timeline
30
 
31
  # Character cards: built-in per-language presets shipped inside the haru package.
32
  _TEMPLATES = card_templates()
 
46
  print("[boot] Haru-4B ready on cuda", flush=True)
47
 
48
 
49
+ @spaces.GPU(duration=25)
50
  def converse(
51
  user_audio,
52
  reference_audio,
 
86
  limit = MAX_USER_SECONDS * SAMPLE_RATE
87
  trimmed = len(user) > limit
88
  user = user[:limit]
89
+ # Full-duplex Haru keeps listening while you speak; a recording that ends
90
+ # mid-sentence cuts the timeline before Haru ever takes its turn. Appending
91
+ # a few seconds of silence gives it the floor, so its reply (and transcript)
92
+ # lands inside the returned timeline.
93
+ user = np.concatenate([user, np.zeros(PAUSE_SECONDS * SAMPLE_RATE, dtype=user.dtype)])
94
  count = max(1, -(-len(user) // frame_samples))
95
  user = np.pad(user, (0, count * frame_samples - len(user)))
96
 
 
140
  text = "".join(frame.text for frame in frames).strip()
141
  if trimmed:
142
  text = f"(input trimmed to {MAX_USER_SECONDS} s)\n{text}"
143
+ if not text:
144
+ text = "… (Haru only listened this time — try again, or raise the temperature.)"
145
  return (SAMPLE_RATE, reply), text
146
 
147
 
requirements.txt CHANGED
@@ -1,2 +1 @@
1
  haru-speech==0.1.0
2
- imageio-ffmpeg
 
1
  haru-speech==0.1.0