Add jitter buffer for discord->ts pipeline

Fixes #1

Signed-off-by: Aron Heinecke <aron.heinecke@t-online.de>

src/discord.rs

@@ -3,6 +3,7 @@
 use serde::Deserialize;
 use serenity::prelude::Mentionable;
 
+use slog::error;
 // This trait adds the `register_songbird` and `register_songbird_with` methods
 // to the client builder below, making it easy to install this voice client.
 // The voice client can be retrieved in any command using `songbird::get(ctx).await`.
@@ -30,8 +31,9 @@ use songbird::{
     EventContext,
     EventHandler as VoiceEventHandler,
 };
+use tsproto_packets::packets::{Direction, InAudioBuf};
 
-use crate::ListenerHolder;
+use crate::{I16_CONVERSION_DIVIDER, ListenerHolder};
 
 pub(crate) struct Handler;
 
@@ -378,24 +380,21 @@ impl VoiceEventHandler for Receiver {
             },
             Ctx::VoicePacket {audio, packet, payload_offset, payload_end_pad} => {
                 // An event which fires for every received audio packet,
-                // containing the decoded data.
+
-                if let Some(audio) = audio {
+                // get raw opus package, we don't decode here and leave that to the AudioHandler
-                    {
+                let last_bytes = packet.payload.len() - payload_end_pad;
-                        let time = std::time::Instant::now();
+                let opus_slice = &packet.payload[*payload_offset..last_bytes];
-                        let mut lock = self.sink.lock().await;
+                let dur;
-                        let dur = time.elapsed();
+                {
-                        if dur.as_millis() > 1 {
+                    let time = std::time::Instant::now();
-                            eprintln!("Acquiring lock took {}ms",dur.as_millis());
+                    let mut lock = self.sink.lock().await;
-                        }
+                    dur = time.elapsed();
-                        if let Some(buffer) = lock.get_mut(&packet.ssrc) {
+                    if let Err(e) = lock.handle_packet(packet.ssrc, packet.sequence.0.0, opus_slice.to_vec()) {
-                            buffer.extend(audio);
+                        eprintln!("Failed to handle Discord voice packet: {}",e);
-                        } else {
-                            // TODO: can we skip this clone ?
-                            let _ = lock.insert(packet.ssrc, audio.clone());
-                        }
                     }
-                } else {
+                }
-                    println!("RTP packet, but no audio. Driver may not be configured to decode.");
+                if dur.as_millis() > 1 {
+                    eprintln!("Acquiring lock took {}ms",dur.as_millis());
                 }
             },
             Ctx::RtcpPacket {packet, payload_offset, payload_end_pad} => {

src/discord_audiohandler.rs

@@ -0,0 +1,498 @@
+//! COPIED FROM tsclientlib https://github.com/ReSpeak/tsclientlib/blob/e4d2baa8aaee5cd793a982e2805d7baf46b715b9/tsclientlib/src/audio.rs
+//! Copyright by their respective owners. 
+//! Adopted to allow usage with non-ts packages.
+//!
+//! Handle receiving audio.
+//!
+//! The [`AudioHandler`] collects all incoming audio packets and queues them per
+//! client. It decodes the audio, handles out-of-order packets and missing
+//! packets. It automatically adjusts the queue length based on the jitter of
+//! incoming packets.
+
+use std::cmp::Reverse;
+use std::collections::{HashMap, VecDeque};
+use std::convert::TryInto;
+use std::fmt::Debug;
+use std::hash::Hash;
+
+use audiopus::coder::Decoder;
+use audiopus::{packet, Channels, SampleRate};
+use slog::{debug, o, trace, warn, Logger};
+use tsclientlib::audio::Error;
+use tsproto_packets::packets::{AudioData, CodecType, InAudioBuf};
+
+use crate::ClientId;
+
+const SAMPLE_RATE: SampleRate = SampleRate::Hz48000;
+const CHANNELS: Channels = Channels::Stereo;
+const CHANNEL_NUM: usize = 2;
+/// If this amount of packets is lost consecutively, we assume the stream stopped.
+const MAX_PACKET_LOSSES: usize = 3;
+/// Store the buffer sizes for the last `LAST_BUFFER_SIZE_COUNT` packets.
+const LAST_BUFFER_SIZE_COUNT: u8 = 255;
+/// The amount of samples to maximally buffer. Equivalent to 0.5 s.
+const MAX_BUFFER_SIZE: usize = 48_000 / 2;
+/// Maximum number of packets in the queue.
+const MAX_BUFFER_PACKETS: usize = 50;
+/// Buffer for maximal 0.5 s without playing anything.
+const MAX_BUFFER_TIME: usize = 48_000 / 2;
+/// Duplicate or remove every `step` sample when speeding-up.
+const SPEED_CHANGE_STEPS: usize = 100;
+/// The usual amount of samples in a frame.
+///
+/// Use 48 kHz, 20 ms frames (50 per second) and mono data (1 channel).
+/// This means 1920 samples and 7.5 kiB.
+const USUAL_FRAME_SIZE: usize = 48000 / 50;
+
+type Result<T> = std::result::Result<T, Error>;
+
+#[derive(Clone, Debug)]
+struct SlidingWindowMinimum<T: Copy + Default + Ord> {
+	/// How long a value stays in the sliding window.
+	size: u8,
+	/// This is a sliding window minimum, it contains
+	/// `(insertion time, value)`.
+	///
+	/// When we insert a value, we can remove all bigger sample counts,
+	/// thus the queue always stays sorted with the minimum at the front
+	/// and the maximum at the back (latest entry).
+	///
+	/// Provides amortized O(1) minimum.
+	/// Source: https://people.cs.uct.ac.za/~ksmith/articles/sliding_window_minimum.html#sliding-window-minimum-algorithm
+	queue: VecDeque<(u8, T)>,
+	/// The current insertion time.
+	cur_time: u8,
+}
+
+#[derive(Debug)]
+struct QueuePacket {
+	packet: Vec<u8>,
+	samples: usize,
+	id: u16,
+}
+
+/// A queue for audio packets for one audio stream.
+pub struct AudioQueue {
+	logger: Logger,
+	decoder: Decoder,
+	pub volume: f32,
+	/// The id of the next packet that should be decoded.
+	///
+	/// Used to check for packet loss.
+	next_id: u16,
+	/// If the last packet was a whisper packet.
+	whispering: bool,
+	packet_buffer: VecDeque<QueuePacket>,
+	/// Amount of samples in the `packet_buffer`.
+	packet_buffer_samples: usize,
+	/// Temporary buffer that contains the samples of one decoded packet.
+	decoded_buffer: Vec<f32>,
+	/// The current position in the `decoded_buffer`.
+	decoded_pos: usize,
+	/// The number of samples in the last packet.
+	last_packet_samples: usize,
+	/// The last `packet_loss_num` packet decodes were a loss.
+	packet_loss_num: usize,
+	/// The amount of samples to buffer until this queue is ready to play.
+	buffering_samples: usize,
+	/// The amount of packets in the buffer when a packet was decoded.
+	///
+	/// Uses the amount of samples in the `packet_buffer` / `USUAL_PACKET_SAMPLES`.
+	/// Used to expand or reduce the buffer.
+	last_buffer_size_min: SlidingWindowMinimum<u8>,
+	last_buffer_size_max: SlidingWindowMinimum<Reverse<u8>>,
+	/// Buffered for this duration.
+	buffered_for_samples: usize,
+}
+
+/// Handles incoming audio, has one [`AudioQueue`] per sending client.
+pub struct AudioHandler<Id: Clone + Debug + Eq + Hash + PartialEq = ClientId> {
+	logger: Logger,
+	queues: HashMap<Id, AudioQueue>,
+	/// Buffer this amount of samples for new queues before starting to play.
+	///
+	/// Updated when a new queue gets added.
+	avg_buffer_samples: usize,
+}
+
+impl<T: Copy + Default + Ord> SlidingWindowMinimum<T> {
+	fn new(size: u8) -> Self { Self { size, queue: Default::default(), cur_time: 0 } }
+
+	fn push(&mut self, value: T) {
+		while self.queue.back().map(|(_, s)| *s >= value).unwrap_or_default() {
+			self.queue.pop_back();
+		}
+		let i = self.cur_time;
+		self.queue.push_back((i, value));
+		while self
+			.queue
+			.front()
+			.map(|(i, _)| self.cur_time.wrapping_sub(*i) >= self.size)
+			.unwrap_or_default()
+		{
+			self.queue.pop_front();
+		}
+		self.cur_time = self.cur_time.wrapping_add(1);
+	}
+
+	fn get_min(&self) -> T { self.queue.front().map(|(_, s)| *s).unwrap_or_default() }
+}
+
+impl AudioQueue {
+	fn new(logger: Logger, sequence: u16, packet: Vec<u8>) -> Result<Self> {
+		let last_packet_samples =
+			packet::nb_samples(&packet, SAMPLE_RATE).map_err(Error::GetPacketSample)?;
+		if last_packet_samples > MAX_BUFFER_SIZE {
+			return Err(Error::TooManySamples);
+		}
+
+		let last_packet_samples = last_packet_samples * CHANNEL_NUM;
+		let mut res = Self {
+			logger,
+			decoder: Decoder::new(SAMPLE_RATE, CHANNELS).map_err(Error::CreateDecoder)?,
+			volume: 1.0,
+			next_id: sequence,
+			whispering: false,
+			packet_buffer: Default::default(),
+			packet_buffer_samples: 0,
+			decoded_buffer: Default::default(),
+			decoded_pos: 0,
+			last_packet_samples,
+			packet_loss_num: 0,
+			buffering_samples: 0,
+			last_buffer_size_min: SlidingWindowMinimum::new(LAST_BUFFER_SIZE_COUNT),
+			last_buffer_size_max: SlidingWindowMinimum::<Reverse<u8>>::new(LAST_BUFFER_SIZE_COUNT),
+			buffered_for_samples: 0,
+		};
+		res.add_buffer_size(0);
+		res.add_packet(sequence, packet)?;
+		Ok(res)
+	}
+
+	pub fn get_decoder(&self) -> &Decoder { &self.decoder }
+	pub fn is_whispering(&self) -> bool { self.whispering }
+
+	/// Size is in samples.
+	fn add_buffer_size(&mut self, size: usize) {
+		if let Ok(size) = (size / USUAL_FRAME_SIZE).try_into() {
+			self.last_buffer_size_min.push(size);
+			self.last_buffer_size_max.push(Reverse(size));
+		} else {
+			warn!(self.logger, "Failed to put amount of packets into an u8"; "size" => size);
+		}
+	}
+
+	/// The approximate deviation of the buffer size.
+	fn get_deviation(&self) -> u8 {
+		let min = self.last_buffer_size_min.get_min();
+		let max = self.last_buffer_size_max.get_min();
+		max.0 - min
+	}
+
+	fn add_packet(&mut self, sequence: u16, packet: Vec<u8>) -> Result<()> {
+		if self.packet_buffer.len() >= MAX_BUFFER_PACKETS {
+			return Err(Error::QueueFull);
+		}
+		let samples;
+		if packet.len() <= 1 {
+			// End of stream
+			samples = 0;
+		} else {
+			samples = packet::nb_samples(&packet, SAMPLE_RATE)
+				.map_err(Error::GetPacketSample)?;
+			if samples > MAX_BUFFER_SIZE {
+				return Err(Error::TooManySamples);
+			}
+		}
+
+		let id = sequence;
+		let packet = QueuePacket { packet, samples, id };
+		if id.wrapping_sub(self.next_id) > MAX_BUFFER_PACKETS as u16 {
+			return Err(Error::TooLate { wanted: self.next_id, got: id });
+		}
+
+		// Put into first spot where the id is smaller
+		let i = self.packet_buffer.len()
+			- self
+				.packet_buffer
+				.iter()
+				.enumerate()
+				.rev()
+				.take_while(|(_, p)| p.id.wrapping_sub(id) <= MAX_BUFFER_PACKETS as u16)
+				.count();
+		// Check for duplicate packet
+		if let Some(p) = self.packet_buffer.get(i) {
+			if p.id == packet.id {
+				return Err(Error::Duplicate(p.id));
+			}
+		}
+
+		trace!(self.logger, "Insert packet {} at {}", id, i);
+		let last_id = self.packet_buffer.back().map(|p| p.id.wrapping_add(1)).unwrap_or(id);
+		if last_id <= id {
+			self.buffering_samples = self.buffering_samples.saturating_sub(samples);
+			// Reduce buffering counter by lost packets if there are some
+			self.buffering_samples = self
+				.buffering_samples
+				.saturating_sub(usize::from(id - last_id) * self.last_packet_samples);
+		}
+
+		self.packet_buffer_samples += packet.samples;
+		self.packet_buffer.insert(i, packet);
+
+		Ok(())
+	}
+
+	fn decode_packet(&mut self, packet: Option<&QueuePacket>, fec: bool) -> Result<()> {
+		trace!(self.logger, "Decoding packet"; "has_packet" => packet.is_some(), "fec" => fec);
+		let packet_data;
+		let len;
+		if let Some(p) = packet {
+			packet_data = Some(&p.packet);
+			len = p.samples;
+			self.whispering = false;
+		} else {
+			packet_data = None;
+			len = self.last_packet_samples;
+		}
+		self.packet_loss_num += 1;
+
+		self.decoded_buffer.resize(self.decoded_pos + len * CHANNEL_NUM, 0.0);
+		let len = self
+			.decoder
+			.decode_float(packet_data.as_deref(), &mut self.decoded_buffer[self.decoded_pos..], fec)
+			.map_err(|e| Error::Decode {
+				error: e,
+				packet: packet.map(|p| p.packet.to_owned()),
+			})?;
+		self.last_packet_samples = len;
+		self.decoded_buffer.truncate(self.decoded_pos + len * CHANNEL_NUM);
+		self.decoded_pos += len * CHANNEL_NUM;
+
+		// Update packet_loss_num
+		if packet.is_some() && !fec {
+			self.packet_loss_num = 0;
+		}
+
+		// Update last_buffer_size
+		let mut count = self.packet_buffer_samples;
+		if let Some(last) = self.packet_buffer.back() {
+			// Lost packets
+			trace!(self.logger, "Ids"; "last_id" => last.id,
+				"next_id" => self.next_id,
+				"first_id" => self.packet_buffer.front().unwrap().id,
+				"buffer_len" => self.packet_buffer.len());
+			count += (usize::from(last.id.wrapping_sub(self.next_id)) + 1
+				- self.packet_buffer.len())
+				* self.last_packet_samples;
+		}
+		self.add_buffer_size(count);
+
+		Ok(())
+	}
+
+	/// Decode data and return the requested length of buffered data.
+	///
+	/// Returns `true` in the second return value when the stream ended,
+	/// `false` when it continues normally.
+	pub fn get_next_data(&mut self, len: usize) -> Result<(&[f32], bool)> {
+		if self.buffering_samples > 0 {
+			if self.buffered_for_samples >= MAX_BUFFER_TIME {
+				self.buffering_samples = 0;
+				self.buffered_for_samples = 0;
+				trace!(self.logger, "Buffered for too long";
+					"buffered_for_samples" => self.buffered_for_samples,
+					"buffering_samples" => self.buffering_samples);
+			} else {
+				self.buffered_for_samples += len;
+				trace!(self.logger, "Buffering";
+					"buffered_for_samples" => self.buffered_for_samples,
+					"buffering_samples" => self.buffering_samples);
+				return Ok((&[], false));
+			}
+		}
+		// Need to refill buffer
+		if self.decoded_pos < self.decoded_buffer.len() {
+			if self.decoded_pos > 0 {
+				self.decoded_buffer.drain(..self.decoded_pos);
+				self.decoded_pos = 0;
+			}
+		} else {
+			self.decoded_buffer.clear();
+			self.decoded_pos = 0;
+		}
+
+		while self.decoded_buffer.len() < len {
+			trace!(self.logger, "get_next_data";
+				"decoded_buffer" => self.decoded_buffer.len(),
+				"decoded_pos" => self.decoded_pos,
+				"len" => len,
+			);
+
+			// Decode a packet
+			if let Some(packet) = self.packet_buffer.pop_front() {
+				if packet.packet.len() <= 1 {
+					// End of stream
+					return Ok((&self.decoded_buffer, true));
+				}
+
+				self.packet_buffer_samples -= packet.samples;
+				let cur_id = self.next_id;
+				self.next_id = self.next_id.wrapping_add(1);
+				if packet.id != cur_id {
+					debug_assert!(
+						packet.id.wrapping_sub(cur_id) < MAX_BUFFER_PACKETS as u16,
+						"Invalid packet queue state: {} < {}",
+						packet.id,
+						cur_id
+					);
+					// Packet loss
+					debug!(self.logger, "Audio packet loss"; "need" => cur_id, "have" => packet.id);
+					if packet.id == self.next_id {
+						// Can use forward-error-correction
+						self.decode_packet(Some(&packet), true)?;
+					} else {
+						self.decode_packet(None, false)?;
+					}
+					self.packet_buffer_samples += packet.samples;
+					self.packet_buffer.push_front(packet);
+				} else {
+					self.decode_packet(Some(&packet), false)?;
+				}
+			} else {
+				debug!(self.logger, "No packets in queue");
+				// Packet loss or end of stream
+				self.decode_packet(None, false)?;
+			}
+
+			if self.last_packet_samples == 0 {
+				break;
+			}
+
+			// Check if we should speed-up playback
+			let min = self.last_buffer_size_min.get_min();
+			let dev = self.get_deviation();
+			if min > (MAX_BUFFER_SIZE / USUAL_FRAME_SIZE) as u8 {
+				debug!(self.logger, "Truncating buffer"; "min" => min);
+				// Throw out all but min samples
+				let mut keep_samples = 0;
+				let keep = self
+					.packet_buffer
+					.iter()
+					.rev()
+					.take_while(|p| {
+						keep_samples += p.samples;
+						keep_samples < usize::from(min) + USUAL_FRAME_SIZE
+					})
+					.count();
+				let len = self.packet_buffer.len() - keep;
+				self.packet_buffer.drain(..len);
+				self.packet_buffer_samples = self.packet_buffer.iter().map(|p| p.samples).sum();
+				if let Some(p) = self.packet_buffer.front() {
+					self.next_id = p.id;
+				}
+			} else if min > dev {
+				// Speed-up
+				debug!(self.logger, "Speed-up buffer"; "min" => min,
+					"cur_packet_count" => self.packet_buffer.len(),
+					"last_packet_samples" => self.last_packet_samples,
+					"dev" => dev);
+				let start = self.decoded_buffer.len() - self.last_packet_samples * CHANNEL_NUM;
+				for i in 0..(self.last_packet_samples / SPEED_CHANGE_STEPS) {
+					let i = start + i * (SPEED_CHANGE_STEPS - 1) * CHANNEL_NUM;
+					self.decoded_buffer.drain(i..(i + CHANNEL_NUM));
+				}
+			}
+		}
+
+		self.decoded_pos = len;
+		Ok((&self.decoded_buffer[..len], false))
+	}
+}
+
+impl<Id: Clone + Debug + Eq + Hash + PartialEq> AudioHandler<Id> {
+	pub fn new(logger: Logger) -> Self {
+		Self { logger, queues: Default::default(), avg_buffer_samples: 0 }
+	}
+
+	/// Delete all queues
+	pub fn reset(&mut self) { self.queues.clear(); }
+
+	pub fn get_queues(&self) -> &HashMap<Id, AudioQueue> { &self.queues }
+	pub fn get_mut_queues(&mut self) -> &mut HashMap<Id, AudioQueue> { &mut self.queues }
+
+	/// `buf` is not cleared before filling it.
+	///
+	/// Returns the clients that are not talking anymore.
+	pub fn fill_buffer(&mut self, buf: &mut [f32]) -> Vec<Id> {
+		self.fill_buffer_with_proc(buf, |_, _| {})
+	}
+
+	/// `buf` is not cleared before filling it.
+	///
+	/// Same as [`fill_buffer`] but before merging a queue into the output buffer, a preprocessor
+	/// function is called. The queue volume is applied after calling the preprocessor.
+	///
+	/// Returns the clients that are not talking anymore.
+	pub fn fill_buffer_with_proc<F: FnMut(&Id, &[f32])>(
+		&mut self, buf: &mut [f32], mut handle: F,
+	) -> Vec<Id> {
+		trace!(self.logger, "Filling audio buffer"; "len" => buf.len());
+		let mut to_remove = Vec::new();
+		for (id, queue) in self.queues.iter_mut() {
+			if queue.packet_loss_num >= MAX_PACKET_LOSSES {
+				debug!(self.logger, "Removing talker"; "packet_loss_num" => queue.packet_loss_num);
+				to_remove.push(id.clone());
+				continue;
+			}
+
+			let vol = queue.volume;
+			match queue.get_next_data(buf.len()) {
+				Err(e) => {
+					warn!(self.logger, "Failed to decode audio packet"; "error" => %e);
+				}
+				Ok((r, is_end)) => {
+					handle(id, &r);
+					for i in 0..r.len() {
+						buf[i] += r[i] * vol;
+					}
+					if is_end {
+						to_remove.push(id.clone());
+					}
+				}
+			}
+		}
+
+		for id in &to_remove {
+			self.queues.remove(&id);
+		}
+		to_remove
+	}
+
+	/// Add a packet to the audio queue.
+	///
+	/// If a new client started talking, returns the id of this client.
+	pub fn handle_packet(&mut self, id: Id, sequence: u16, packet: Vec<u8>) -> Result<Option<Id>> {
+		if let Some(queue) = self.queues.get_mut(&id) {
+			queue.add_packet(sequence, packet)?;
+			Ok(None)
+		} else {
+
+			trace!(self.logger, "Adding talker");
+			let mut queue =
+				AudioQueue::new(self.logger.new(o!("client" => format!("{:?}", id))), sequence,packet)?;
+			if !self.queues.is_empty() {
+				// Update avg_buffer_samples
+				self.avg_buffer_samples = USUAL_FRAME_SIZE
+					+ self
+						.queues
+						.values()
+						.map(|q| usize::from(q.last_buffer_size_min.get_min()))
+						.sum::<usize>() / self.queues.len();
+			}
+			queue.buffering_samples = self.avg_buffer_samples;
+			self.queues.insert(id.clone(), queue);
+			Ok(Some(id))
+		}
+	}
+}

src/main.rs

@@ -11,6 +11,7 @@ use tokio::sync::Mutex;
 use anyhow::*;
 
 mod discord;
+mod discord_audiohandler;
 
 #[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
 struct ConnectionId(u64);
@@ -47,7 +48,7 @@ struct Config {
 struct ListenerHolder;
 
 //TODO: stop shooting myself in the knee with a mutex
-type AudioBufferDiscord = Arc<Mutex<HashMap<u32,Vec<i16>>>>;
+type AudioBufferDiscord = Arc<Mutex<discord_audiohandler::AudioHandler<u32>>>;
 
 
 type TsVoiceId = (ConnectionId, ClientId);
@@ -93,7 +94,12 @@ impl TypeMapKey for ListenerHolder {
 /// We want to run every 20ms, but we only get ~1ms correctness
 const TICK_TIME: u64 = 18;
 const FRAME_SIZE_MS: usize = 20;
-const STEREO_20MS: usize = 48000 * 2 * FRAME_SIZE_MS / 1000;
+const SAMPLE_RATE: usize = 48000;
+const STEREO_20MS: usize = SAMPLE_RATE * 2 * FRAME_SIZE_MS / 1000;
+// const STEREO_20MS_FLOAT: usize = SAMPLE_RATE / 20;
+/// See http://blog.bjornroche.com/2009/12/int-float-int-its-jungle-out-there.html
+/// We use i16::MIN here, which is 0x8000
+const I16_CONVERSION_DIVIDER: f32 = 0x8000 as f32;
 /// The maximum size of an opus frame is 1275 as from RFC6716.
 const MAX_OPUS_FRAME_SIZE: usize = 1275;
 #[tokio::main]
@@ -137,8 +143,8 @@ async fn main() -> Result<()> {
 	let teamspeak_voice_handler = TsToDiscordPipeline::new(ts_voice_logger);
 
 	// init discord -> teamspeak pipeline
-	let map = HashMap::new();
+	let discord_voice_logger = logger.new(o!("pipeline" => "voice-discord"));
-	let discord_voice_buffer: AudioBufferDiscord = Arc::new(Mutex::new(map));
+	let discord_voice_buffer: AudioBufferDiscord = Arc::new(Mutex::new(discord_audiohandler::AudioHandler::new(discord_voice_logger)));
 	// stuff discord -> teamspeak pipeline into discord context for retrieval inside the client
 	{
 		// Open the data lock in write mode, so keys can be inserted to it.
@@ -206,7 +212,7 @@ async fn main() -> Result<()> {
 				let mut ts_voice: std::sync::MutexGuard<TsAudioHandler> = teamspeak_voice_handler.data.lock().expect("Can't lock ts audio buffer!");
 				// feed mixer+jitter buffer, consumed by discord
 				if let Err(e) = ts_voice.handle_packet((con_id, from), packet) {
-					debug!(logger, "Failed to play TS_Voice packet"; "error" => %e);
+					debug!(logger, "Failed to handle TS_Voice packet"; "error" => %e);
 				}
 			}
 			Ok(())
@@ -243,37 +249,26 @@ async fn main() -> Result<()> {
 /// Create an audio frame for consumption by teamspeak.
 /// Merges all streams and converts them to opus
 async fn process_discord_audio(voice_buffer: &AudioBufferDiscord, encoder: &Arc<Mutex<Encoder>>) -> Option<OutPacket> {
-	let mut buffer_map;
+	// let mut buffer_map;
+	// {
+	// 	let mut lock = voice_buffer.lock().await;
+	// 	buffer_map = std::mem::replace(&mut *lock, HashMap::new());
+	// }
+
+	let mut data = [0.0; STEREO_20MS];
 	{
 		let mut lock = voice_buffer.lock().await;
-		buffer_map = std::mem::replace(&mut *lock, HashMap::new());
+		lock.fill_buffer(&mut data);
-	}
-	if buffer_map.is_empty() {
-		return None;
 	}
 	let mut encoded = [0; 1024];
 	let encoder_c = encoder.clone();
 	// don't block the async runtime
 	let res = task::spawn_blocking(move || {
-		let start = std::time::Instant::now();
+		let start = std::time::Instant::now();		
-		let mut data: Vec<i16> = Vec::with_capacity(STEREO_20MS);
-		// merge all audio buffers (clients) to one
-		for buffer in buffer_map.values_mut() {
-			//buffer.truncate(STEREO_20MS);
-			for i in 0..buffer.len() {
-				if let Some(v) = data.get_mut(i) {
-					*v = *v + buffer[i];
-				} else {
-					data.extend(&buffer[i..]);
-					break;
-				}
-			}
-		}
-		
 		// encode back to opus
 		// this should never block, thus we don't fail gracefully for it
 		let lock = encoder_c.try_lock().expect("Can't reach encoder!");
-		let length = match lock.encode(&data, &mut encoded) {
+		let length = match lock.encode_float(&data, &mut encoded) {
 			Err(e) => {eprintln!("Failed to encode voice: {}",e); return None;},
 			Ok(size) => size,
 		};
@@ -281,7 +276,7 @@ async fn process_discord_audio(voice_buffer: &AudioBufferDiscord, encoder: &Arc<
 		//println!("length size: {}",length);
 		// warn on high encoding times
 		let duration = start.elapsed().as_millis();
-		if duration > 5 {
+		if duration > 2 {
 			eprintln!("Took too {}ms for processing audio!",duration);
 		}
 		// package into teamspeak audio structure