QOL updates

README.md

@@ -1,6 +1,6 @@
 ---
 title: YouTube Streaming ASR
-emoji: 💩
+emoji: 📺
 colorFrom: red
 colorTo: red
 sdk: streamlit

app.py

@@ -19,6 +19,8 @@ player_options = {
     "config": {"youtube": {"playerVars": {"start": 1}}},
 }
 
+# disable rapid fading in and out on `st.code` updates
+st.markdown("<style>.element-container{opacity:1 !important}</style>", unsafe_allow_html=True)
 
 @st.cache(hash_funcs={torch.nn.parameter.Parameter: lambda _: None})
 def load_model(model_path="facebook/wav2vec2-large-robust-ft-swbd-300h"):
@@ -75,7 +77,7 @@ def main():
         state.pad_duration_ms = st.slider("Padding duration (ms)", 100, 5000, 1000, 100)
         submit_button = st.form_submit_button(label="Submit")
 
-    if submit_button or "asr_stream" not in state:
+    if submit_button:
         # a hack to update the video player on value changes
         state.youtube_url = (
             state.youtube_url.split("&hash=")[0]
@@ -85,7 +87,7 @@ def main():
             state.youtube_url, state.chunk_duration_ms, state.pad_duration_ms
         )
         state.chunks_taken = 0
-        state.lines = deque([], maxlen=3)  # limit to the last 3 lines of subs
+        state.lines = deque([], maxlen=5)  # limit to the last 5 lines of subs
 
     player = st_player(state.youtube_url, **player_options, key="youtube_player")
 

streaming.py

@@ -41,15 +41,13 @@ def ffmpeg_stream(youtube_url, sampling_rate=16_000, chunk_duration_ms=5000, pad
     except FileNotFoundError:
         raise ValueError("ffmpeg was not found but is required to stream audio files from filename")
 
-    running = True
     acc = b""
     leftover = np.zeros((0,), dtype=np.float32)
-    while running and ytdl_process.poll() is None:
+    while ytdl_process.poll() is None:
         buflen = read_chunk_len * size_of_sample
 
         raw = ffmpeg_process.stdout.read(buflen)
         if raw == b"":
-            running = False
             break
 
         if len(acc) + len(raw) > buflen:
@@ -61,7 +59,6 @@ def ffmpeg_stream(youtube_url, sampling_rate=16_000, chunk_duration_ms=5000, pad
         audio = np.concatenate([leftover, audio])
         if len(audio) < pad_len * 2:
             # TODO: handle end of stream better than this
-            running = False
             break
         yield audio