malarinv
/
jasper-asr
mirror of https://github.com/malarinv/jasper-asr.git
1
0
Fork 0
Code Issues Releases Wiki Activity

1. added option to strip silent chunks 

2. computing caller quality based on task-id of corrections
Malar Kannan 2020-06-17 21:42:20 +05:30
parent ac0e04c226
commit 000853b600
2 changed files with 21 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
jasper/data/call_recycler.py
View File

@ -121,6 +121,7 @@ def analyze(
), ),
start_delay: float = 1.5, start_delay: float = 1.5,
download_only: bool = False, download_only: bool = False,
strip_silent_chunks: bool = True,
call_logs_file: Path = typer.Option(Path("./call_logs.yaml"), show_default=True), call_logs_file: Path = typer.Option(Path("./call_logs.yaml"), show_default=True),
output_dir: Path = Path("./data"), output_dir: Path = Path("./data"),
data_name: str = None, data_name: str = None,
@ -408,6 +409,16 @@ def analyze(
pprint(call_plots) pprint(call_plots)
def extract_data_points(): def extract_data_points():
if strip_silent_chunks:
def audio_process(seg):
return strip_silence(seg)
else:
def audio_process(seg):
return seg
def gen_data_values(saved_wav_path, data_points, caller_name): def gen_data_values(saved_wav_path, data_points, caller_name):
call_seg = ( call_seg = (
AudioSegment.from_wav(saved_wav_path) AudioSegment.from_wav(saved_wav_path)
@ -417,12 +428,15 @@ def analyze(
) )
for dp_id, dp in enumerate(data_points): for dp_id, dp in enumerate(data_points):
start, end, spoken = dp["start_time"], dp["end_time"], dp["code"] start, end, spoken = dp["start_time"], dp["end_time"], dp["code"]
spoken_seg = strip_silence(call_seg[start * 1000 : end * 1000]) spoken_seg = audio_process(call_seg[start * 1000 : end * 1000])
spoken_fb = BytesIO() spoken_fb = BytesIO()
spoken_seg.export(spoken_fb, format="wav") spoken_seg.export(spoken_fb, format="wav")
spoken_wav = spoken_fb.getvalue() spoken_wav = spoken_fb.getvalue()
# search for actual pnr code and handle plain codes as well # search for actual pnr code and handle plain codes as well
extracted_code = text_extractor(spoken) extracted_code = text_extractor(spoken)
if strip_silent_chunks and spoken_seg.duration_seconds < 0.5:
print(f'transcript chunk "{spoken}" contains no audio skipping.')
continue
yield extracted_code, spoken_seg.duration_seconds, spoken_wav, caller_name, spoken_seg yield extracted_code, spoken_seg.duration_seconds, spoken_wav, caller_name, spoken_seg
call_lens = lens["users"].Each()["calls"].Each() call_lens = lens["users"].Each()["calls"].Each()

7
jasper/data/validation/process.py
View File

@ -194,6 +194,7 @@ def dump_corrections(
@app.command() @app.command()
def caller_quality( def caller_quality(
task_uid: str,
data_name: str = typer.Option("call_upwork_train_cnd", show_default=True), data_name: str = typer.Option("call_upwork_train_cnd", show_default=True),
dump_dir: Path = Path("./data/asr_data"), dump_dir: Path = Path("./data/asr_data"),
dump_fname: Path = Path("ui_dump.json"), dump_fname: Path = Path("ui_dump.json"),
@ -214,7 +215,11 @@ def caller_quality(
dp["valid"] = c["value"]["status"] == "Correct" dp["valid"] = c["value"]["status"] == "Correct"
return dp return dp
corrected_dump = [correction_dp(c) for c in correction_data] corrected_dump = [
correction_dp(c)
for c in correction_data
if c["task_id"].rsplit("-", 1)[1] == task_uid
]
df = pd.DataFrame(corrected_dump) df = pd.DataFrame(corrected_dump)
print(f"Total samples: {len(df)}") print(f"Total samples: {len(df)}")
for (c, g) in df.groupby("caller"): for (c, g) in df.groupby("caller"):