diff --git a/docs/source/conf.py b/docs/source/conf.py
index 599df8b3e..88522ff27 100644
--- a/docs/source/conf.py
+++ b/docs/source/conf.py
@@ -33,6 +33,7 @@ release = "0.1"
 # ones.
 extensions = [
     "sphinx_rtd_theme",
+    "sphinx.ext.todo",
 ]
 
 # Add any paths that contain templates here, relative to this directory.
@@ -74,3 +75,5 @@ html_context = {
     "github_version": "master",
     "conf_py_path": "/icefall/docs/source/",
 }
+
+todo_include_todos = True
diff --git a/docs/source/recipes/aishell.rst b/docs/source/recipes/aishell.rst
deleted file mode 100644
index 71ccaa1fc..000000000
--- a/docs/source/recipes/aishell.rst
+++ /dev/null
@@ -1,10 +0,0 @@
-Aishell
-=======
-
-We provide the following models for the Aishell dataset:
-
-.. toctree::
-   :maxdepth: 2
-
-   aishell/conformer_ctc
-   aishell/tdnn_lstm_ctc
diff --git a/docs/source/recipes/aishell/index.rst b/docs/source/recipes/aishell/index.rst
new file mode 100644
index 000000000..d072d6e9c
--- /dev/null
+++ b/docs/source/recipes/aishell/index.rst
@@ -0,0 +1,22 @@
+aishell
+=======
+
+Aishell is an open-source Chinese Mandarin speech corpus published by Beijing
+Shell Shell Technology Co.,Ltd.
+
+400 people from different accent areas in China are invited to participate in
+the recording, which is conducted in a quiet indoor environment using high
+fidelity microphone and downsampled to 16kHz. The manual transcription accuracy
+is above 95%, through professional speech annotation and strict quality
+inspection. The data is free for academic use. We hope to provide moderate
+amount of data for new researchers in the field of speech recognition.
+
+It can be downloaded from `<https://www.openslr.org/33/>`_
+
+.. toctree::
+   :maxdepth: 1
+
+   tdnn_lstm_ctc
+   conformer_ctc
+   stateless_transducer
+
diff --git a/docs/source/recipes/aishell/stateless_transducer.rst b/docs/source/recipes/aishell/stateless_transducer.rst
new file mode 100644
index 000000000..46b52b257
--- /dev/null
+++ b/docs/source/recipes/aishell/stateless_transducer.rst
@@ -0,0 +1,221 @@
+Stateless Transducer
+====================
+
+This tutorial shows you how to do transducer training in ``icefall``.
+
+.. HINT::
+
+  Instead of using RNN-T or RNN transducer, we only use transducer
+  here. As you will see, there are no RNNs in the model.
+
+The Model
+---------
+
+The transducer model consists of 3 parts:
+
+- **Encoder**: It is a conformer encoder with the following parameters
+
+    - Number of heads: 8
+    - Attention dim: 512
+    - Number of layers: 12
+    - Feedforward dim: 2048
+
+- **Decoder**: We use a stateless model consisting of:
+
+    - An embedding layer with embedding dim 512
+    - A Conv1d layer with a default kernel size 2
+
+- **Joiner**: It consists of a ``nn.tanh()`` and a ``nn.Linear()``.
+
+.. Caution::
+
+  The decoder is stateless and very simple. It is borrowed from
+  `<https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419>`_
+  (Rnn-Transducer with Stateless Prediction Network)
+
+  We make one modification to it: Place a Conv1d layer right after
+  the embedding layer.
+
+When using Chinese characters as modelling unit, whose vocabulary size
+is 4335 in this specific dataset,
+the number of parameters of the model is ``87939824``, i.e., about ``88 M``.
+
+The Loss
+--------
+
+We are using `<https://github.com/csukuangfj/optimized_transducer>`_
+to compute the transducer loss, which removes extra paddings
+in loss computation to save memory.
+
+.. Hint::
+
+  ``optimized_transducer`` implements the technqiues proposed
+  in `Improving RNN Transducer Modeling for End-to-End Speech Recognition <https://arxiv.org/abs/1909.12415>`_ to save memory.
+
+  Furthermore, it supports ``modified transducer``, limiting the maximum
+  number of symbols that can be emitted per frame to 1, which simplifies
+  the decoding process significantly. Also, the experiment results
+  show that it does not degrade the performance.
+
+  See `<https://github.com/csukuangfj/optimized_transducer#modified-transducer>`_
+  for what exactly modified transducer is.
+
+  `<https://github.com/csukuangfj/transducer-loss-benchmarking>`_ shows that
+  in the unpruned case ``optimized_transducer`` has the advantage about minimizing
+  memory usage.
+
+.. todo::
+
+  Add tutorial about ``pruned_transducer_stateless`` that uses k2
+  pruned transducer loss.
+
+.. hint::
+
+  You can use::
+
+    pip install optimized_transducer
+
+  to install ``optimized_transducer``. Refer to
+  `<https://github.com/csukuangfj/optimized_transducer>`_ for other
+  alternatives.
+
+Data Preparation
+----------------
+
+To prepare the data for training, please use the following commands:
+
+.. code-block:: bash
+
+  cd egs/aishell/ASR
+  ./prepare.sh --stop-stage 4
+  ./prepare.sh --stage 6 --stop-stage 6
+
+.. note::
+
+  You can use ``./prepare.sh``, though it will generates FSTs that
+  are not used in transducer traning.
+
+When you finish running the script, you will get the following two folders:
+
+  - ``data/fbank``: It saves the pre-computed features
+  - ``data/lang_char``: It contains tokens that will be used in the training
+
+Training
+--------
+
+.. code-block:: bash
+
+  cd egs/aishell/ASR
+  ./transducer_stateless_modified/train.py --help
+
+shows you the training options that can be passed from the commandline.
+The following options are used quite often:
+
+  - ``--exp-dir``
+
+    The experiment folder to save logs and model checkpoints,
+    defaults to ``./transducer_stateless_modified/exp``.
+
+  - ``--num-epochs``
+
+    It is the number of epochs to train. For instance,
+    ``./transducer_stateless_modified/train.py --num-epochs 30`` trains for 30
+    epochs and generates ``epoch-0.pt``, ``epoch-1.pt``, ..., ``epoch-29.pt``
+    in the folder set by ``--exp-dir``.
+
+  - ``--start-epoch``
+
+    It's used to resume training.
+    ``./transducer_stateless_modified/train.py --start-epoch 10`` loads the
+    checkpoint from ``exp_dir/epoch-9.pt`` and starts
+    training from epoch 10, based on the state from epoch 9.
+
+  - ``--world-size``
+
+    It is used for multi-GPU single-machine DDP training.
+
+      - (a) If it is 1, then no DDP training is used.
+
+      - (b) If it is 2, then GPU 0 and GPU 1 are used for DDP training.
+
+    The following shows some use cases with it.
+
+      **Use case 1**: You have 4 GPUs, but you only want to use GPU 0 and
+      GPU 2 for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="0,2"
+          $ ./transducer_stateless_modified/train.py --world-size 2
+
+      **Use case 2**: You have 4 GPUs and you want to use all of them
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ ./transducer_stateless_modified/train.py --world-size 4
+
+      **Use case 3**: You have 4 GPUs but you only want to use GPU 3
+      for training. You can do the following:
+
+        .. code-block:: bash
+
+          $ cd egs/aishell/ASR
+          $ export CUDA_VISIBLE_DEVICES="3"
+          $ ./transducer_stateless_modified/train.py --world-size 1
+
+    .. CAUTION::
+
+      Only multi-GPU single-machine DDP training is implemented at present.
+      There is an on-going PR `<https://github.com/k2-fsa/icefall/pull/63>`
+      that adds support for multi-GPU multi-machine DDP training.
+
+  - ``--max-duration``
+
+    It specifies the number of seconds over all utterances in a
+    batch, before **padding**.
+    If you encounter CUDA OOM, please reduce it. For instance, if
+    your are using V100 NVIDIA GPU with 32 GB RAM, we recommend you
+    to set it to ``300``.
+
+    .. HINT::
+
+      Due to padding, the number of seconds of all utterances in a
+      batch will usually be larger than ``--max-duration``.
+
+      A larger value for ``--max-duration`` may cause OOM during training,
+      while a smaller value may increase the training time. You have to
+      tune it.
+
+  - ``--lr-factor``
+
+    It contrals the learning rate. If you use single GPU training, you
+    may want to use a small value for it. If you use multiple GPUs for training,
+    you may increase it.
+
+  - ``--context-size``
+
+    It specifies the kernel size in the decoder. Default value 2 means it
+    functions as a tri-gram LM.
+
+  - ``--modified-transducer-prob``
+
+    It specifies the probability to use modified transducer loss.
+    If it is 0, then no modified transducer is used; if it is 1,
+    then it uses modified transducer loss for all batches. If it is
+    ``p``, it applies modified transducer with probability ``p``.
+
+There are some training options, e.g.,
+number of warmup steps,
+that are not passed from the commandline.
+They are pre-configured by the function ``get_params()`` in
+`transducer_stateless_modified/train.py <https://github.com/k2-fsa/icefall/blob/master/egs/aishell/ASR/transducer_stateless_modified/train.py#L162>`_
+
+If you need to change them, please modify ``./transducer_stateless_modified/train.py`` directly.
+
+.. CAUTION::
+
+  The training set is perturbed by speed with two factors: 0.9 and 1.1.
+  Each epoch actually processes ``3x150 == 450`` hours of data.
diff --git a/docs/source/recipes/index.rst b/docs/source/recipes/index.rst
index 78e9ea569..9d1d83d29 100644
--- a/docs/source/recipes/index.rst
+++ b/docs/source/recipes/index.rst
@@ -10,12 +10,10 @@ We may add recipes for other tasks as well in the future.
 .. Other recipes are listed in a alphabetical order.
 
 .. toctree::
-   :maxdepth: 3
+   :maxdepth: 2
+   :caption: Table of Contents
 
-   yesno
-
-   librispeech
-
-   aishell
-
-   timit
+   aishell/index
+   librispeech/index
+   timit/index
+   yesno/index
diff --git a/docs/source/recipes/librispeech.rst b/docs/source/recipes/librispeech.rst
deleted file mode 100644
index 946b23407..000000000
--- a/docs/source/recipes/librispeech.rst
+++ /dev/null
@@ -1,10 +0,0 @@
-LibriSpeech
-===========
-
-We provide the following models for the LibriSpeech dataset:
-
-.. toctree::
-   :maxdepth: 2
-
-   librispeech/tdnn_lstm_ctc
-   librispeech/conformer_ctc
diff --git a/docs/source/recipes/librispeech/index.rst b/docs/source/recipes/librispeech/index.rst
new file mode 100644
index 000000000..5fa08ab6b
--- /dev/null
+++ b/docs/source/recipes/librispeech/index.rst
@@ -0,0 +1,8 @@
+LibriSpeech
+===========
+
+.. toctree::
+   :maxdepth: 1
+
+   tdnn_lstm_ctc
+   conformer_ctc
diff --git a/docs/source/recipes/timit.rst b/docs/source/recipes/timit.rst
deleted file mode 100644
index b630e2ce4..000000000
--- a/docs/source/recipes/timit.rst
+++ /dev/null
@@ -1,10 +0,0 @@
-TIMIT
-===========
-
-We provide the following models for the TIMIT dataset:
-
-.. toctree::
-   :maxdepth: 2
-
-   timit/tdnn_lstm_ctc
-   timit/tdnn_ligru_ctc
\ No newline at end of file
diff --git a/docs/source/recipes/timit/index.rst b/docs/source/recipes/timit/index.rst
new file mode 100644
index 000000000..17f40cdb7
--- /dev/null
+++ b/docs/source/recipes/timit/index.rst
@@ -0,0 +1,9 @@
+TIMIT
+=====
+
+.. toctree::
+   :maxdepth: 1
+
+   tdnn_ligru_ctc
+   tdnn_lstm_ctc
+
diff --git a/docs/source/recipes/timit/tdnn_ligru_ctc.rst b/docs/source/recipes/timit/tdnn_ligru_ctc.rst
index 30877505f..186420ee7 100644
--- a/docs/source/recipes/timit/tdnn_ligru_ctc.rst
+++ b/docs/source/recipes/timit/tdnn_ligru_ctc.rst
@@ -1,5 +1,5 @@
 TDNN-LiGRU-CTC
-=============
+==============
 
 This tutorial shows you how to run a TDNN-LiGRU-CTC model with the `TIMIT <https://data.deepai.org/timit.zip>`_ dataset.
 
diff --git a/docs/source/recipes/images/yesno-tdnn-tensorboard-log.png b/docs/source/recipes/yesno/images/tdnn-tensorboard-log.png
similarity index 100%
rename from docs/source/recipes/images/yesno-tdnn-tensorboard-log.png
rename to docs/source/recipes/yesno/images/tdnn-tensorboard-log.png
diff --git a/docs/source/recipes/yesno/index.rst b/docs/source/recipes/yesno/index.rst
new file mode 100644
index 000000000..d68523a97
--- /dev/null
+++ b/docs/source/recipes/yesno/index.rst
@@ -0,0 +1,7 @@
+YesNo
+=====
+
+.. toctree::
+   :maxdepth: 1
+
+   tdnn
diff --git a/docs/source/recipes/yesno.rst b/docs/source/recipes/yesno/tdnn.rst
similarity index 99%
rename from docs/source/recipes/yesno.rst
rename to docs/source/recipes/yesno/tdnn.rst
index cb425ad1d..e8b748e6b 100644
--- a/docs/source/recipes/yesno.rst
+++ b/docs/source/recipes/yesno/tdnn.rst
@@ -1,5 +1,5 @@
-yesno
-=====
+TDNN-CTC
+========
 
 This page shows you how to run the `yesno <https://www.openslr.org/1>`_ recipe. It contains:
 
@@ -145,7 +145,7 @@ In ``tdnn/exp``, you will find the following files:
     Note there is a URL in the above output, click it and you will see
     the following screenshot:
 
-      .. figure:: images/yesno-tdnn-tensorboard-log.png
+      .. figure:: images/tdnn-tensorboard-log.png
          :width: 600
          :alt: TensorBoard screenshot
          :align: center