Lyra is a high-quality, low-bitrate speech codec that makes voice communication available even on the slowest networks. To do this it applies traditional codec techniques while leveraging advances in machine learning (ML) with models trained on thousands of hours of data to create a novel method for compressing and transmitting voice signals.

The basic architecture of the Lyra codec is quite simple. Features are extracted from speech every 40ms and are then compressed for transmission at a bitrate of 3kbps. The features themselves are log mel spectrograms, a list of numbers representing the speech energy in different frequency bands, which have traditionally been used for their perceptual relevance because they are modeled after human auditory response. On the other end, a generative model uses those features to recreate the speech signal.

Lyra harnesses the power of new natural-sounding generative models to maintain the low bitrate of parametric codecs while achieving high quality, on par with state-of-the-art waveform codecs used in most streaming and communication platforms today.

Computational complexity is reduced by using a cheaper recurrent generative model, a WaveRNN variation, that works at a lower rate, but generates in parallel multiple signals in different frequency ranges that it later combines into a single output signal at the desired sample rate. This trick, plus 64-bit ARM optimizations, enables Lyra to not only run on cloud servers, but also on-device on mid-range phones, such as Pixel phones, in real time (with a processing latency of 100ms). This generative model is then trained on thousands of hours of speech data with speakers in over 70 languages and optimized to accurately recreate the input audio.

There are a few things you'll need to do to set up your computer to build Lyra.

Lyra is built using Google's build system, Bazel. Install it following these instructions. Bazel verson 4.0.0 is required, and some Linux distributions may make an older version available in their application repositories, so make sure you are using the required version or newer. The latest version can be downloaded via Github.

Lyra can be built from linux using bazel for an arm android target, or a linux target. The android target is optimized for realtime performance. The linux target is typically used for development and debugging.

Building on android requires downloading a specific version of the android NDK toolchain. If you develop with Android Studio already, you might not need to do these steps if ANDROID_HOME and ANDROID_NDK_HOME are defined and pointing at the right version of the NDK.

1. Download the sdk manager from https://developer.android.com/studio
2. Unzip and cd to the directory
3. Check the available packages to install in case they don't match the following steps.

bin/sdkmanager  --sdk_root=$HOME/android/sdk --list

Some systems will already have the java runtime set up. But if you see an error here like ERROR: JAVA_HOME is not set and no 'java' command could be found on your PATH., this means you need to install the java runtime with sudo apt install default-jdk first. You will also need to add export JAVA_HOME=/usr/lib/jvm/java-11-openjdk-amd64 (type ls /usr/lib/jvm to see which path was installed) to your $HOME/.bashrc and reload it with source $HOME/.bashrc.

4. Install the r21 ndk, android sdk 29, and build tools:

bin/sdkmanager  --sdk_root=$HOME/android/sdk --install  "platforms;android-29" "build-tools;29.0.3" "ndk;21.4.7075529"

5. Add the following to .bashrc (or export the variables)

export ANDROID_NDK_HOME=$HOME/android/sdk/ndk/21.4.7075529
export ANDROID_HOME=$HOME/android/sdk

6. Reload .bashrc (with source $HOME/.bashrc)

The building and running process differs slightly depending on the selected platform.

You can build the cc_binaries with the default config. encoder_main is an example of a file encoder.

bazel build -c opt :encoder_main

You can run encoder_main to encode a test .wav file with some speech in it, specified by --input_path. The --model_path flag contains the model data necessary to encode, and --output_path specifies where to write the encoded (compressed) representation.

bazel-bin/encoder_main --model_path=wavegru --output_dir=$HOME/temp --input_path=testdata/16khz_sample_000001.wav

Similarly, you can build decoder_main and use it on the output of encoder_main to decode the encoded data back into speech.

bazel build -c opt :decoder_main
bazel-bin/decoder_main  --model_path=wavegru --output_dir=$HOME/temp/ --encoded_path=$HOME/temp/16khz_sample_000001.lyra

Note: the default Bazel toolchain is automatically configured and likely uses gcc/libstdc++ on Linux. This should be satisfactory for most users, but will differ from the NDK toolchain, which uses clang/libc++. To use a custom clang toolchain on Linux, see toolchain/README.md and .bazelrc.

There is an example APK target called lyra_android_example that you can build after you have set up the NDK.

This example is an app with a minimal GUI that has buttons for two options. One option is to record from the microphone and encode/decode with Lyra so you can test what Lyra would sound like for your voice. The other option runs a benchmark that encodes and decodes in the background and prints the timings to logcat.

bazel build android_example:lyra_android_example --config=android_arm64 --copt=-DBENCHMARK
adb install bazel-bin/android_example/lyra_android_example.apk

After this you should see an app called "Lyra Example App".

You can open it, and you will see a simple TextView that says the benchmark is running, and when it finishes.

Press "Record from microphone", say a few words (be sure to have your microphone near your mouth), and then press "Encode and decode to speaker". You should hear your voice being played back after being coded with Lyra.

If you press 'Benchmark', you should see something like the following in logcat on a Pixel 4 when running the benchmark:

I  Starting benchmarkDecode()
I  I20210401 11:04:06.898649  6870 lyra_wavegru.h:75] lyra_wavegru running fast multiplication kernels for aarch64.
I  I20210401 11:04:06.900411  6870 layer_wrapper.h:162] |lyra_16khz_ar_to_gates_| layer:  Shape: [3072, 4]. Sparsity: 0
I  I20210401 11:04:07.031975  6870 layer_wrapper.h:162] |lyra_16khz_gru_layer_| layer:  Shape: [3072, 1024]. Sparsity: 0.9375
...
I  I20210401 11:04:26.700160  6870 benchmark_decode_lib.cc:167] Using float arithmetic.
I  I20210401 11:04:26.700352  6870 benchmark_decode_lib.cc:85] conditioning_only stats for generating 2000 frames of audio, max: 506 us, min: 368 us, mean: 391 us, stdev: 10.3923.
I  I20210401 11:04:26.725538  6870 benchmark_decode_lib.cc:85] model_only stats for generating 2000 frames of audio, max: 12690 us, min: 9087 us, mean: 9237 us, stdev: 262.416.
I  I20210401 11:04:26.729460  6870 benchmark_decode_lib.cc:85] combined_model_and_conditioning stats for generating 2000 frames of audio, max: 13173 us, min: 9463 us, mean: 9629 us, stdev: 270.788.
I  Finished benchmarkDecode()

This shows that decoding a 25Hz frame (each frame is .04 seconds) takes 9629 microseconds on average (.0096 seconds). So decoding is performed at around 4.15 (.04/.0096) times faster than realtime.

For even faster decoding, you can use a fixed point representation by building with --copt=-DUSE_FIXED16, although there may be some loss of quality.

To build your own android app, you can either use the cc_library target outputs to create a .so that you can use in your own build system. Or you can use it with an android_binary rule within bazel to create an .apk file as in this example.

There is a tutorial on building for android with Bazel in the bazel docs.

There are also the binary targets that you can use to experiment with encoding and decoding .wav files.

You can build the example cc_binary targets with:

bazel build -c opt :encoder_main --config=android_arm64
bazel build -c opt :decoder_main --config=android_arm64

This builds an executable binary that can be run on android 64-bit arm devices (not an android app). You can then push it to your android device and run it as a binary through the shell.

# Push the binary and the data it needs, including the model and .wav files:
adb push bazel-bin/encoder_main /data/local/tmp/
adb push bazel-bin/decoder_main /data/local/tmp/
adb push wavegru/ /data/local/tmp/
adb push testdata/ /data/local/tmp/

adb shell
cd /data/local/tmp
./encoder_main --model_path=/data/local/tmp/wavegru --output_dir=/data/local/tmp --input_path=testdata/16khz_sample_000001.wav
./decoder_main --model_path=/data/local/tmp/wavegru --output_dir=/data/local/tmp --encoded_path=16khz_sample_000001.lyra

The encoder_main/decoder_main as above should also work.

For integrating Lyra into any project only two APIs are relevant: LyraEncoder and LyraDecoder.

DISCLAIMER: At this time Lyra's API and bit-stream are not guaranteed to be stable and might change in future versions of the code.

On the sending side, LyraEncoder can be used to encode an audio stream using the following interface:

class LyraEncoder : public LyraEncoderInterface {
 public:
  static std::unique_ptr<LyraEncoder> Create(
      int sample_rate_hz, int num_channels, int bitrate, bool enable_dtx,
      const ghc::filesystem::path& model_path);

  absl::optional<std::vector<uint8_t>> Encode(
      const absl::Span<const int16_t> audio) override;

  int sample_rate_hz() const override;

  int num_channels() const override;

  int bitrate() const override;

  int frame_rate() const override;
};

The static Create method instantiates a LyraEncoder with the desired sample rate in Hertz, number of channels and bitrate, as long as those parameters are supported. Else it returns a nullptr. The Create method also needs to know if DTX should be enabled and where the model weights are stored. It also checks that these weights exist and are compatible with the current Lyra version.

Given a LyraEncoder, any audio stream can be compressed using the Encode method. The provided span of int16-formatted samples is assumed to contain 40ms of data at the sample rate chosen at Create time. As long as this condition is met the Encode method returns the encoded packet as a vector of bytes that is ready to be stored or transmitted over the network.

The rest of the LyraEncoder methods are just getters for the different predetermined parameters.

On the receiving end, LyraDecoder can be used to decode the encoded packet using the following interface:

class LyraDecoder : public LyraDecoderInterface {
 public:
  static std::unique_ptr<LyraDecoder> Create(
      int sample_rate_hz, int num_channels, int bitrate,
      const ghc::filesystem::path& model_path);

  bool SetEncodedPacket(absl::Span<const uint8_t> encoded) override;

  absl::optional<std::vector<int16_t>> DecodeSamples(int num_samples) override;

  absl::optional<std::vector<int16_t>> DecodePacketLoss(
      int num_samples) override;

  int sample_rate_hz() const override;

  int num_channels() const override;

  int bitrate() const override;

  int frame_rate() const override;

  bool is_comfort_noise() const override;
};

Once again, the static Create method instantiates a LyraDecoder with the desired sample rate in Hertz, number of channels and bitrate, as long as those parameters are supported. Else it returns a nullptr. These parameters don't need to be the same as the ones in LyraEncoder. And once again, the Create method also needs to know where the model weights are stored. It also checks that these weights exist and are compatible with the current Lyra version.

Given a LyraDecoder, any packet can be decoded by first feeding it into SetEncodedPacket, which returns true if the provided span of bytes is a valid Lyra-encoded packet.

Then the int16-formatted samples can be obtained by calling DecodeSamples, as long as the total number of samples obtained this way between any two calls to SetEncodedPacket is less than 40ms of data at the sample rate chose at Create time.

If there isn't a packet available, but samples still need to be generated, DecodePacketLoss can be used, which doesn't have a restriction on the number of samples.

In those cases, the decoder might switch to a comfort noise generation mode, which can be checked using is_confort_noise.

The rest of the LyraDecoder methods are just getters for the different predetermined parameters.

For an example on how to use LyraEncoder and LyraDecoder to encode and decode a stream of audio, please refer to the integration test.

Lyra uses a library in the sparse_matmul directory that enables fast execution of sparse Matrix-Vector multiplication ops on mobile and desktop CPU platforms (ARM and AVX2) to allow for real-time operation on phones. This library was created by DeepMind for their implementation of WaveRNN with sparsity [4], which gave a huge improvement in complexity over WaveNet.

A generic kernel is also provided, which enables debugging on non-optimized platforms. Contributions for other platforms are welcome.

Use of this source code is governed by a Apache v2.0 license that can be found in the LICENSE file.

1. Kleijn, W. B., Lim, F. S., Luebs, A., Skoglund, J., Stimberg, F., Wang, Q., & Walters, T. C. (2018, April). Wavenet based low rate speech coding. In 2018 IEEE international conference on acoustics, speech and signal processing (ICASSP) (pp. 676-680). IEEE.
2. Denton, T., Luebs, A., Lim, F. S., Storus, A., Yeh, H., Kleijn, W. B., & Skoglund, J. (2021). Handling Background Noise in Neural Speech Generation. arXiv preprint arXiv:2102.11906.
3. Kleijn, W. B., Storus, A., Chinen, M., Denton, T., Lim, F. S., Luebs, A., ... & Yeh, H. (2021). Generative Speech Coding with Predictive Variance Regularization. arXiv preprint arXiv:2102.09660. 4. Kalchbrenner, N., Elsen, E., Simonyan, K., Noury, S., Casagrande, N., Lockhart, E., ... & Kavukcuoglu, K. (2018, July). Efficient neural audio synthesis. In International Conference on Machine Learning (pp. 2410-2419). PMLR.