diff --git a/examples/scripts/asr/nnet3-keep-loglikes.py b/examples/scripts/asr/nnet3-keep-loglikes.py
new file mode 100755
index 000000000..ce551d39d
--- /dev/null
+++ b/examples/scripts/asr/nnet3-keep-loglikes.py
@@ -0,0 +1,80 @@
+#!/usr/bin/env python
+
+## This script is very similar to the second part in ./nnet3-online-recognizer.py,
+## but it has additional code to extract the log_likelihoods from the nnet
+## during decoding.  Instead of dumping to stdout, the numpy arrays could be saved
+## to disc for later recognition using a script similar to ./mapped-loglikes-recognizer.py. 
+
+from __future__ import print_function
+
+from kaldi.asr import NnetLatticeFasterOnlineRecognizer
+from kaldi.decoder import LatticeFasterDecoderOptions
+from kaldi.nnet3 import NnetSimpleLoopedComputationOptions
+from kaldi.online2 import (OnlineEndpointConfig,
+                           OnlineIvectorExtractorAdaptationState,
+                           OnlineNnetFeaturePipelineConfig,
+                           OnlineNnetFeaturePipelineInfo,
+                           OnlineNnetFeaturePipeline,
+                           OnlineSilenceWeighting)
+from kaldi.util.options import ParseOptions
+from kaldi.util.table import SequentialWaveReader
+
+chunk_size = 1440
+
+# Define online feature pipeline
+feat_opts = OnlineNnetFeaturePipelineConfig()
+endpoint_opts = OnlineEndpointConfig()
+po = ParseOptions("")
+feat_opts.register(po)
+endpoint_opts.register(po)
+po.read_config_file("online.conf")
+feat_info = OnlineNnetFeaturePipelineInfo.from_config(feat_opts)
+
+# Construct recognizer
+decoder_opts = LatticeFasterDecoderOptions()
+decoder_opts.beam = 23
+decoder_opts.max_active = 7000
+decodable_opts = NnetSimpleLoopedComputationOptions()
+decodable_opts.acoustic_scale = 1.0
+decodable_opts.frame_subsampling_factor = 3
+decodable_opts.frames_per_chunk = 50 ## smallish to force many updates
+asr = NnetLatticeFasterOnlineRecognizer.from_files(
+    "final.mdl", "HCLG.fst", "words.txt",
+    decoder_opts=decoder_opts,
+    decodable_opts=decodable_opts,
+    endpoint_opts=endpoint_opts)
+
+# Decode (chunked + partial output + log_likelihoods)
+for key, wav in SequentialWaveReader("scp:wav.scp"):
+    feat_pipeline = OnlineNnetFeaturePipeline(feat_info)
+    asr.set_input_pipeline(feat_pipeline)
+    d = asr._decodable
+    asr.init_decoding()
+    data = wav.data()[0]
+    last_chunk = False
+    part = 1
+    prev_num_frames_decoded = 0
+    prev_num_frames_computed = 0
+    for i in range(0, len(data), chunk_size):
+        if i + chunk_size >= len(data):
+            last_chunk = True
+        feat_pipeline.accept_waveform(wav.samp_freq, data[i:i + chunk_size])
+        if last_chunk:
+            feat_pipeline.input_finished()
+        nr = d.num_frames_ready()
+        if nr > prev_num_frames_computed:
+            x = d.log_likelihoods(prev_num_frames_computed, nr).numpy()
+            print(x.shape, x)
+            prev_num_frames_computed = nr
+        asr.advance_decoding()
+        num_frames_decoded = asr.decoder.num_frames_decoded()
+        if not last_chunk:
+            if num_frames_decoded > prev_num_frames_decoded:
+                prev_num_frames_decoded = num_frames_decoded
+                out = asr.get_partial_output()
+                print(key + "-part%d" % part, out["text"], flush=True)
+                part += 1
+    asr.finalize_decoding()
+    out = asr.get_output()
+    print(key + "-final", out["text"], flush=True)
+
diff --git a/kaldi/nnet3/decodable-online-looped.clif b/kaldi/nnet3/decodable-online-looped.clif
index f0fe8925e..d08740e28 100644
--- a/kaldi/nnet3/decodable-online-looped.clif
+++ b/kaldi/nnet3/decodable-online-looped.clif
@@ -2,8 +2,10 @@ from "itf/online-feature-itf-clifwrap.h" import *
 from "hmm/transition-model-clifwrap.h" import *
 from "nnet3/decodable-simple-looped-clifwrap.h" import *
 from "matrix/kaldi-vector-clifwrap.h" import *
+from "matrix/kaldi-matrix-clifwrap.h" import *
 
 from kaldi.matrix._matrix import _vector_wrapper
+from kaldi.matrix._matrix import _matrix_wrapper
 from kaldi.itf._decodable_itf import DecodableInterface
 
 from "nnet3/decodable-online-looped.h":
@@ -29,7 +31,7 @@ from "nnet3/decodable-online-looped.h":
       def `FrameSubsamplingFactor` as frame_subsampling_factor(self) -> int:
         """Returns the frame subsampling factor."""
 
-      def `LogLikelihoods` as log_likelihoods(self, frame: int) -> Vector:
+      def `LogLikelihoods` as log_likelihoods_frame(self, frame: int) -> Vector:
         """Returns the log-likelihoods for the given frame"""
         return _vector_wrapper(...)
 
@@ -55,7 +57,11 @@ from "nnet3/decodable-online-looped.h":
       def `FrameSubsamplingFactor` as frame_subsampling_factor(self) -> int:
         """Returns the frame subsampling factor."""
 
-      def `LogLikelihoods` as log_likelihoods(self, frame: int) -> Vector:
+      def `LogLikelihoods` as log_likelihoods_frame(self, frame: int) -> Vector:
         """Returns the log-likelihoods for the given frame"""
 	return _vector_wrapper(...)
 
+      def `LogLikelihoods` as log_likelihoods(self, frame_from: int, frame_to: int) -> Matrix:
+        """Returns the log-likelihoods for the given frame"""
+	return _matrix_wrapper(...)
+