gotgcall

README

gotgcall — Pure-Go Telegram Group Call & Voice Chat Streaming Library

gotgcall is a pure-Go library for streaming audio and video into Telegram group calls (voice chats and video chats). It is a drop-in alternative to ntgcalls / pytgcalls built for Go music bots, livestream bots, and broadcast tooling.

Use it to:

• Build a Telegram music bot in Go that joins a voice chat and plays MP3/FLAC/M4A/Opus/HLS or any audio.
• Stream a live video broadcast (mp4/mkv/webm/RTMP/RTSP) into a Telegram group call.
• Push a "go live" RTMP broadcast to a channel via phone.GetGroupCallStreamRtmpUrl.
• Wrap any ffmpeg pipeline as a streaming source — atempo, scaling, hardware encoders, atomic source switches.

No libwebrtc, no cgo, no native build chain. CGO_ENABLED=0 go build produces a single static binary on every supported platform. WebRTC runs on pion v4; ffmpeg is invoked as a runtime binary for transcoding only — nothing is linked in.

Keywords: Telegram group call · Telegram voice chat · pure Go WebRTC · ntgcalls Go alternative · pytgcalls Go · pion WebRTC Telegram · Telegram music bot Go · gogram voice chat · Telegram video chat streaming · RTMP push Telegram livestream

Status

Work in progress. Built for my own bots; the API is intentionally close to ntgcalls so existing code translates with minimal change.

Contents

• Install
• Architecture at a glance
• Quick start
• Sources
  • FromFile / FromURL
  • FromShell — single custom ffmpeg leg
  • FromShells — dual ffmpeg legs
  • EncodeOptions
• Client options
• Lifecycle
  • WebRTC mode
  • RTMP mode
• Pause / Resume / Mute
• Callbacks
• Server-side media-state changes
• Errors
• Concurrency model
• Networking
• Performance notes
• A/V sync
• Pitfalls
• Performance vs ntgcalls
• Why pure Go
• License

Install

go get github.com/annihilatorrrr/gotgcall

ffmpeg must be on PATH at runtime (or set gotgcall.WithFFmpegPath("/path/to/ffmpeg")). New() fails fast if the binary isn't found, so the error surfaces at startup rather than on the first stream.

Requires Go 1.26+ (uses errors.AsType[T] and a few stdlib features added in 1.26).

Architecture at a glance

                       ┌──────────────────────────────┐
                       │           Client             │  one process-wide handle
                       │  (gotgcall.go)               │  multiplexes any number of calls
                       └──────────────────────────────┘
                            │           │
                            ▼           ▼
                   ┌──────────────┐  ┌──────────────┐
                   │  GroupCall   │  │   RTMPCall   │  per-chat call instance
                   │  (WebRTC)    │  │  (FFmpeg→RTMP│
                   └──────────────┘  └──────────────┘
                            │              │
                            │              └── single ffmpeg push to Telegram's RTMP URL
                            ▼
                   ┌──────────────┐   ┌──────────────┐
                   │   Streamer   │──▶│ pion Track   │──▶ Telegram SFU
                   │ (paces opus/ │   │ Local Static │
                   │  ivf frames) │   │   Sample     │
                   └──────────────┘   └──────────────┘
                            ▲
                            │ media.Sample (Opus / VP8)
                            │
                   ┌──────────────┐   ┌──────────────┐
                   │ FrameReader  │◀──│ ShellReader  │◀── ffmpeg subprocess
                   │ (OGG / IVF)  │   │ (stdout pipe)│
                   └──────────────┘   └──────────────┘

Blob-only signaling. The library never imports gogram or any MTProto layer. CreateCall(chatID) returns a JSON string; the caller passes it to phone.JoinGroupCall via their own MTProto stack, then hands the response back via Connect(chatID, respJSON). This keeps the library MTProto-version-independent.

One PeerConnection per call. Send-only audio (Opus PT=111) and video (VP8 PT=100). All calls share one wrtc.Factory (and optionally one UDP socket; see WithSharedUDPMux).

ffmpeg outputs ENCODED Opus (OGG) and VP8 (IVF), not raw PCM/YUV. Pion's TrackLocalStaticSample.WriteSample expects already-encoded frames, so we let ffmpeg do the encoding and skip a Go-side Opus encoder (which would force cgo). This also saves ~48× pipe bandwidth versus PCM.

Quick start

client, err := gotgcall.New()
if err != nil { log.Fatal(err) }
defer client.Close()

client.OnStreamEnd(func(chat int64, t gotgcall.StreamType, d gotgcall.Device, err error) {
    log.Printf("stream end: %v", err)
})
client.OnConnectionChange(func(chat int64, info gotgcall.NetworkInfo) {
    log.Printf("conn state: %s", info.State)
})
client.OnMediaStateChange(func(chat int64, state gotgcall.MediaState) {
    // Mirror to phone.EditGroupCallParticipant so Telegram knows the
    // bot just toggled video/mute/pause. Required for /play → /vplay
    // to actually show video to other participants.
})

// 1. Local-side JSON.
localParams, _ := client.CreateCall(chatID)

// 2. Drive Telegram via your MTProto layer (gogram, etc.).
//    Pass localParams to phone.JoinGroupCall; read the response.
remoteParams := joinViaYourMTProto(localParams)

// 3. Finish the WebRTC handshake.
client.Connect(chatID, remoteParams)

// 4. Stream.
client.SetStreamSources(chatID, gotgcall.FromFile("song.mp3", gotgcall.EncodeOptions{}))

// 5. Pause / resume / mute / change source any time.
client.Pause(chatID)
client.Resume(chatID)
client.SetStreamSources(chatID, gotgcall.FromURL("https://stream.example.com/radio.m3u8", gotgcall.EncodeOptions{}))

// 6. Stop tears down the call.
client.Stop(chatID)

See examples/bot/ for a runnable skeleton against gogram (own go.mod so the example doesn't taint the library's dependency tree).

Sources

All sources target Opus-in-OGG (audio) and/or VP8-in-IVF (video) on ffmpeg's stdout. The library will not accept raw PCM/YUV — the frame readers can't parse them.

FromFile / FromURL

gotgcall.FromFile("song.mp3", gotgcall.EncodeOptions{})
gotgcall.FromURL("https://stream.example.com/...", gotgcall.EncodeOptions{})

Anything ffmpeg can decode is fair game — mp3, m4a, flac, ogg, opus, wav, webm, mp4, mkv, mov, m3u8 (HLS), live RTMP/RTSP, etc.

Defaults to audio only, regardless of what the container holds. Opt in to video extraction:

client.SetStreamSources(chatID, gotgcall.FromFile("movie.mp4", gotgcall.EncodeOptions{
    Tracks: gotgcall.TrackAudio | gotgcall.TrackVideo,
    // Or just TrackVideo — TrackVideo implies TrackAudio (a video file is a
    // video file with audio).
}))

Fast-start probing (-analyzeduration 0 -probesize 64k) is on by default for every source — cuts ~1-2 s off ffmpeg's startup latency vs the stock defaults (5 s + 5 MB). HLS sources additionally get -user_agent, -protocol_whitelist file,http,https,tcp,tls, -rw_timeout 10s, -http_persistent 1; HTTP/HTTPS sources get -reconnect 1 -reconnect_at_eof 1 -reconnect_streamed 1 -reconnect_delay_max 5 -timeout 10s so transient network blips don't kill the stream.

Both FromFile and FromURL return seekable sources. Pause records the elapsed offset and Resume re-spawns ffmpeg with -ss <offset> injected before the input.

FromShell — single custom ffmpeg leg

gotgcall.FromShell("ffmpeg -i thing.mp3", gotgcall.TrackAudio)

FromShell parses the cmdline as a shell-like argv (handles double-quoted args, plus \" and \\ escape sequences for filenames containing literal " or \ — e.g. a Telegram audio titled (From "Foo") that would otherwise slice the path mid-string when the embedded quote toggled the quote state) and spawns it directly via exec, NOT via /bin/sh. Shell metacharacters in filenames can't inject commands; use %q for filenames!

Missing essentials are filled in automatically:

• Input-side (always on): fast-start probing + -err_detect ignore_err before -i.
• Output-side (audio): -c:a libopus, -application audio, -frame_duration 20, -page_duration 20000, -mapping_family 0, -ar 48000, -ac 2, -f ogg, pipe:1.
• Output-side (video): -c:v libvpx, -deadline realtime, -f ivf, pipe:1.

So the minimum command works:

gotgcall.FromShell(`ffmpeg -i "song.mp3"`, gotgcall.TrackAudio)

…and is equivalent to the fully-spelled-out form:

gotgcall.FromShell(`ffmpeg -analyzeduration 0 -probesize 64k -err_detect ignore_err `+
    `-i "song.mp3" -vn -c:a libopus -b:a 64k -application audio `+
    `-frame_duration 20 -page_duration 20000 -mapping_family 0 `+
    `-ar 48000 -ac 2 -f ogg pipe:1`, gotgcall.TrackAudio)

Video-only example:

gotgcall.FromShell(`ffmpeg -i "movie.mp4" -an -c:v libvpx -deadline realtime `+
    `-b:v 800k -vf scale=1280:720 -r 30 -f ivf pipe:1`, gotgcall.TrackVideo)

A single FromShell call produces a single output (audio OR video). Raw PCM/YUV output codecs (-c:a pcm_*, -f rawvideo, etc.) are rejected up front with a useful error.

FromShells — dual ffmpeg legs

For ntgcalls-style "microphone + camera" patterns where you want full control over both legs:

gotgcall.FromShells(
    `ffmpeg -i "x.mp4"`,                                            // audio leg
    `ffmpeg -i "x.mp4" -vf scale=1280:720 -b:v 1500k`,              // video leg
)

Each cmd goes through the same auto-flag injection as FromShell. Either string may be empty to skip that track.

For the convenience path use FromFile/FromURL with Tracks: TrackVideo and let the library construct both ffmpeg commands for you.

EncodeOptions

type EncodeOptions struct {
    VideoBitrateKbps int   // default 800
    VideoWidth       int   // default 1280
    VideoHeight      int   // default 720
    VideoFPS         int   // default 30
    AudioBitrateKbps int   // default 64
    AudioChannels    int   // default 2
    Tracks           Track // default TrackAudio; TrackVideo implies +TrackAudio
}

Set on the constructor (FromFile/FromURL); rides with the Source. FromShell / FromShells ignore EncodeOptions because you control ffmpeg directly.

Client options

gotgcall.New(
    gotgcall.WithFFmpegPath("/opt/ffmpeg/bin/ffmpeg"),  // override binary lookup
    gotgcall.WithLogger(slog.Default()),                // structured logger
    gotgcall.WithDebugLogs(),                           // shortcut: text handler @ Debug level to stderr
    gotgcall.WithFFmpegStderrLog(),                     // tee ffmpeg stderr → debug log
    gotgcall.WithSharedUDPMux(),                        // one UDP socket for all calls
    gotgcall.WithDTLSCertPool(16),                      // pre-generate N DTLS certs
    gotgcall.WithDispatchBuffer(512),                   // event-dispatcher queue size
    gotgcall.WithICEServers([]gotgcall.ICEServer{       // STUN + TURN
        {URLs: []string{"stun:stun.l.google.com:19302"}},
        {URLs: []string{"turn:turn.example.com:3478"},
         Username: "u", Credential: "p"},
    }),
    gotgcall.WithNetworkTypes(                          // enable IPv6/TCP for restrictive nets
        gotgcall.NetworkTypeUDP4,
        gotgcall.NetworkTypeUDP6,
        gotgcall.NetworkTypeTCP4,
    ),
    gotgcall.WithICETimeouts(60*time.Second, 120*time.Second, 5*time.Second),
)

Option	Default	Notes
WithFFmpegPath	"ffmpeg"	New() fails fast with exec.LookPath if the binary is missing.
WithLogger	discard	Plumbed into the WebRTC factory, the media package (ffmpeg stderr/exit), the dispatcher, and pion's internal ICE/DTLS/SCTP logs via the slog bridge in wrtc/pion_logger.go.
WithDebugLogs	off	Convenience shortcut that installs a slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{Level: slog.LevelDebug}). Use when reporting bugs.
WithFFmpegStderrLog	off	Tees ffmpeg stderr line-by-line into the logger at Debug level. Without this, stderr is only surfaced in the final error message (last 512 bytes) when ffmpeg crashes — useless for "stream runs but I hear nothing" symptoms.
WithSharedUDPMux	off	Opens one udp4:0 socket and routes ICE for every call through it. See UDP mux scaling below.
WithDTLSCertPool	8	Background goroutine keeps N pre-generated ECDSA-P256 certs ready so CreateCall doesn't stall on keygen during bursts. 0 = disabled.
WithDispatchBuffer	256	Size of the single callback-dispatcher channel. Larger absorbs bursts of state changes before the consumer drains.
WithICEServers	2× Google STUN	Overrides the default STUN list. Add TURN entries for users behind symmetric NAT / restrictive firewalls. Pass an empty slice to disable STUN entirely (host-only candidates).
WithNetworkTypes	UDP4+UDP6	Override the ICE candidate network-type whitelist. Add TCP for restrictive environments where UDP is blocked.
WithICETimeouts	60 s / 120 s / 2 s	(disconnect, failed, keepalive). Generous defaults because Telegram's edge wobble on rejoin frequently takes 60-90 s to settle on a working candidate pair. Pass 0 for any value to keep the default; ultra-responsive UIs can shorten back to 30/60.

Enabling debug logs

If you've heard "I set WithLogger but I see nothing" — before the slog bridge below was added, pion's internal logs (ICE state, DTLS handshake, SCTP) went straight to stderr via the log package, completely bypassing WithLogger. That is now fixed: the bridge wraps every pion logger into your slog handler, tagged with pion=<scope> (e.g. pion=ice, pion=dtls).

The fastest way to see everything:

client, err := gotgcall.New(
    gotgcall.WithDebugLogs(),
    gotgcall.WithFFmpegStderrLog(),
)

With both on, you get: gotgcall internals (Debug), pion ICE/DTLS/SCTP/interceptor (Debug+), ffmpeg stderr lines (Debug). Filter by attribute key if it's too much:

slog.SetDefault(slog.New(slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{
    Level: slog.LevelDebug,
    ReplaceAttr: func(_ []string, a slog.Attr) slog.Attr {
        // example: drop pion=interceptor lines
        if a.Key == "pion" && a.Value.String() == "interceptor" {
            return slog.Attr{}
        }
        return a
    },
})))

UDP mux & scaling

The old README said "use WithSharedUDPMux at 100+ calls". That was a conservative guess — the real picture:

Default (one socket per call):

• 1 UDP socket = 1 file descriptor + 1 ephemeral port per call.
• Linux defaults: ulimit -n 1024 (raise to 65535), ephemeral port range 32768–60999 (~28000 usable).
• Practical ceiling without any tuning: ~900 calls (bounded by FDs, leaving room for other FDs).
• After ulimit -n 65535 and net.ipv4.ip_local_port_range="1024 65000": tens of thousands of calls on a beefy server.
• Benefit: kernel-level UDP receive-queue per call, parallelism scales with CPU cores naturally.

WithSharedUDPMux (one socket total):

• 1 UDP socket, 1 FD, 1 port for the entire process — FD/port limits stop mattering.
• All traffic funnels through one socket → kernel UDP buffer might become contended at extreme rates.
• Per-socket UDP throughput on modern Linux: easily 1–10 Gbps. At ~50 kbps per voice call, that's 20 000–200 000 concurrent voice calls through one socket before throughput becomes the bottleneck.
• Best for huge call counts where FD/port pressure is the limiting factor, or where firewall rules need to pin a single port.

Rule of thumb:

• < 1000 calls: per-call sockets is fine, simpler, and gives you natural per-call kernel-queue isolation.
• 1000–10000 calls: either works; WithSharedUDPMux simplifies sysctl tuning.
• 10000+ calls: WithSharedUDPMux is the easier path; tune the kernel UDP receive buffer (net.core.rmem_max, net.core.rmem_default).

Note: client.Stop(chatID) closes only that call's PeerConnection (and the per-call socket if not using the shared mux). The shared mux survives every Stop and is only closed when you call client.Close() on the parent client. So you can spin calls up and down freely without leaking or thrashing the shared socket.

Lifecycle

WebRTC mode

The default. Use for normal group voice/video.

localParams, err := client.CreateCall(chatID)
// → send localParams to phone.JoinGroupCall; read remoteParams from response.
err = client.Connect(chatID, remoteParams)
err = client.SetStreamSources(chatID, gotgcall.FromFile("song.mp3", gotgcall.EncodeOptions{}))
// …
err = client.Stop(chatID)

• CreateCall returns ErrConnectionExists if a call for that chat already exists. Per-chat creation mutex serialises concurrent calls so it never allocates twice.
• Connect is idempotent only in the sense that re-calling it re-SetRemoteDescription's; if you call Connect before CreateCall you get ErrConnectionNotFound.
• Stop removes the call from the internal map and clears the per-chat mutex; after Stop you can re-use the same chatID cleanly.
• client.AudioSSRC(chatID) returns the audio SSRC for phone.LeaveGroupCall's Source field. RTMP calls return ErrWrongMode.

RTMP mode

For "go live" / host-style broadcasts. Obtain the URL via phone.GetGroupCallStreamRtmpUrl:

err := client.StartRTMP(chatID, rtmpURL)
err  = client.SetStreamSources(chatID, gotgcall.FromFile("movie.mp4", gotgcall.EncodeOptions{}))
// Pause/Resume/Stop work identically. Mute/Unmute are best-effort (RTMP push has
// no per-track control); the lib tracks state but doesn't drop frames.

StartRTMP is serialised with CreateCall via the same per-chat mutex. RTMP transcodes to H.264+AAC and pushes FLV.

Pause in RTMP mode is kill-and-restart-with--ss (Telegram's RTMP ingest times out silent streams, so SIGSTOP can't be used). WebRTC mode uses a channel-based gate that keeps ffmpeg alive — the OS pipe absorbs ~1s of frames during pause.

Pause / Resume / Mute

ok, err := client.Pause(chatID)   // false if already paused
ok, err  = client.Resume(chatID)
ok, err  = client.Mute(chatID)    // mute audio track; video keeps going
ok, err  = client.Unmute(chatID)

WebRTC mode:

• Pause gate-blocks the streamer's pull loop. ffmpeg keeps running; its stdout pipe buffers the next ~1s of frames. Resume wakes the loop and the pacing baseline jumps forward over the paused window so we don't burst the buffered frames on resume.
• Mute is a flag on the streamer — samples are read at the natural cadence but WriteSample is skipped.

RTMP mode:

• Pause records elapsed_ms, kills ffmpeg, frees the connection. Resume spawns a fresh ffmpeg with -ss <elapsed>.

SetStreamSources can be called any time. While paused the new source is recorded but not started; Resume starts it at offset 0 (a new source resets the resume offset — it's a different track).

Callbacks

client.OnStreamEnd(func(chat int64, t StreamType, d Device, err error) {
    // Fires on natural EOF, ffmpeg crash, Stop. err == nil for clean EOF/Stop.
})

client.OnConnectionChange(func(chat int64, info NetworkInfo) {
    // info.State: Connecting | Connected | Disconnected | Failed | Closed | Timeout
})

client.OnMediaStateChange(func(chat int64, state MediaState) {
    // Fires on every Muted / Paused / VideoStopped transition. Wire to
    // your MTProto layer's phone.EditGroupCallParticipant so Telegram
    // mirrors the change for other participants. Critical for the
    // /play → /vplay swap: without flipping VideoStopped=false on the
    // server side, Telegram's SFU silently drops the late video even
    // though our RTP is correct.
})

All callbacks fire on a single dispatcher goroutine, so you can safely re-enter the API from inside (e.g. call client.Stop(chat) from inside OnStreamEnd). If your callback panics it is recovered and logged; the dispatcher keeps running.

If the dispatch queue fills up (slow consumer), the dispatcher drops the oldest queued event and logs a warning. Tune with WithDispatchBuffer.

Server-side media-state changes (admin mute, video off)

The library is blob-only and never sees MTProto updates. When Telegram tells you the bot was admin-muted (via your UpdateGroupCallParticipants handler), react directly:

tg.AddRawHandler(&telegram.UpdateGroupCallParticipants{}, func(u telegram.Update, _ *telegram.Client) error {
    upd := u.(*telegram.UpdateGroupCallParticipants)
    for _, p := range upd.Participants {
        // compare p.Peer to your own user id, then:
        if p.Muted {
            client.Pause(chatID)
        } else if p.CanSelfUnmute {
            client.Resume(chatID)
        }
    }
    return nil
})

There is no OnUpgrade / NotifyUpgrade API by design — out of scope for a blob-only library.

Errors

All errors are sentinels — branch with errors.Is:

Error	Returned when
ErrConnectionExists	CreateCall/StartRTMP for a chatID that already has a live call.
ErrConnectionNotFound	Any method called with an unknown chatID, or after Stop.
ErrConnectionTimeout	Declared for future use (currently surfaced via OnConnectionChange(Failed) after pion's 120 s ICE-failed timeout).
ErrConnectionFailed	Same — declared for branching; current ICE-failed manifests as OnConnectionChange(Failed).
ErrInvalidParams	Malformed remote JSON in Connect (missing ufrag/pwd/fingerprints), or FromShell with empty/invalid command.
ErrFFmpegSpawn	exec.Cmd.Start failed (binary missing / permission denied / OS resource exhaustion).
ErrFFmpegCrashed	ffmpeg exited non-zero; wrapped error carries exit=<code> and the last 512 bytes of stderr for diagnosis. Surfaced both via OnStreamEnd and on the ShellReader.Read EOF path (the Reader briefly waits — bounded 200 ms — for the reap goroutine to capture the real exit before returning, so a fast-failing child no longer collapses to a bare io.EOF swallowed by the OGG/IVF parser).
ErrFile	Source contained no playable audio or video stream (OGG / IVF parse failed).
ErrClosed	Any method called after Client.Close().
ErrNotConnected	SetSource timed out waiting for WebRTC to reach Connected (15 s).
ErrInternal	Wrapping for pion API errors that shouldn't happen (e.g. CreateOffer failure).
ErrWrongMode	WebRTC-only method called on an RTMP call (or vice versa).

Concurrency model

• One *Client per process multiplexes any number of group calls.
• All public methods are safe for concurrent use.
• Per-chat operations are serialised internally via a sync.RWMutex on each call instance.
• Concurrent CreateCall / StartRTMP for the same chat are gated by a per-chat creation mutex; the first wins, others get ErrConnectionExists without allocating a pion PeerConnection.
• The createMu map entry is freed in Stop (you can re-use the chatID cleanly).
• Callbacks fire on a single dispatcher goroutine — no inter-callback parallelism, but no risk of deadlocking the producer either.

Networking

Transport. Pion v4 is the only WebRTC stack. Full ICE mode with no STUN servers — we gather only host candidates (fast, no STUN round-trips) but still actively send connectivity checks to Telegram's SFU. This matches ntgcalls' actual wire behavior (its ICEMODE_LITE declaration is dead code; it runs full ICE). UDP4+UDP6 enabled by default (matching ntgcalls' PORTALLOCATOR_ENABLE_IPV6). DTLS role is passive/server (matching ntgcalls' SSL_SERVER).

Interface filter. Virtual / VPN interfaces are skipped by name match: vethernet, vmware, virtualbox, vbox, hyper-v, loopback, teredo, isatap, tap-, docker, wsl, tailscale, zerotier, openvpn. Gathering candidates on these would slow ICE and produce unreachable pairs.

STUN. Two Google STUN servers are configured by default so pion can gather server-reflexive candidates behind NAT. Override with WithICEServers; pass an empty slice to disable STUN entirely (host-only candidates — fine when the bot has a public IP). TURN is not configured by default; pass WithICEServers with TURN entries for users behind symmetric NAT or restrictive firewalls.

ICE-lite answer. The synthesized answer SDP carries a=ice-lite at session level, telling pion that Telegram's SFU is ICE-lite (server-side, never sends connectivity checks). This lets pion's ICE state machine skip waiting for reverse checks and nominate pairs faster — without it, intermittent connection timeouts can occur when pion's timing heuristics expect checks from the remote that never arrive.

ICE timeouts. Disconnect grace = 60 s, failed declaration = 120 s, keepalive = 2 s. Generous defaults because Telegram's edge wobble on rejoin takes 60-90 s to settle on a working candidate pair. Override via WithICETimeouts for ultra-responsive UIs (shorter) or extra-unstable networks (longer). Pion surfaces failure via OnConnectionStateChange(Failed). On top of the pion timers the FactoryMonitor runs a 30 s checking-stuck safety net: if a PC stays in Connecting for more than 30 s the monitor force-closes it. This is a last resort — the SetSource connection gate (15 s) fires first with a clean ErrNotConnected for normal use.

UDP mux. Default behavior: each call binds its own UDP socket. Enable WithSharedUDPMux() to route every call through one shared udp4:0 socket. Useful at 100+ concurrent calls where you don't want N ephemeral ports open.

RTP header extensions. The full Telegram-required set is registered: ssrc-audio-level (RFC 6464), abs-send-time, transport-cc, sdes-mid, video-orientation. The library auto-stamps ssrc-audio-level (-20 dBov, voice-activity bit set) and abs-send-time on every outbound audio packet via a pion interceptor — Telegram's SFU silently drops streams that don't carry audio-level (it treats them as silence and stops forwarding to listeners).

Outbound RTP marker bit. Pion's packetizer sets marker=true on every single-payload Opus packet, but per RFC 7587 the marker should only be set on the first packet after silence. An always-set marker forces jitter-buffer resync at the SFU and degrades audio. We clear it via a small interceptor on outbound audio.

Pion log noise filter. Telegram's mixer forwards every other participant's RTP to us; our PeerConnection has only send-only tracks, so pion logs Simulcast probing failed for each unknown incoming SSRC. We filter these out at the Error level so they don't bury real errors. Other levels pass through.

HLS / HTTP. ffmpeg-side, not pion-side. See FromFile / FromURL for the auto-injected reconnect / timeout flags.

Performance notes

• Cert pool. ECDSA-P256 keygen is ~10ms per call. The cert pool keeps N ready so burst joins don't queue behind keygen latency. Defaults to 8; raise for very bursty workloads.
• Single dispatcher. All callbacks serialise on one goroutine. Tune WithDispatchBuffer if you see drop warnings.
• Single timer per streamer. The pacing loop reuses one time.Timer for the whole stream rather than allocating a NewTimer per sample (Go 1.23+ Reset semantics make this safe without manual drain).
• OS-pipe-managed stdout. ShellReader uses os.Pipe rather than cmd.StdoutPipe so cmd.Wait doesn't close the read end out from under us. Without this, the last chunk of audio buffered in the kernel pipe would be discarded the moment ffmpeg exits.
• Per-page OGG flush. -page_duration 20000 on libopus forces ffmpeg to emit one OGG page per Opus frame. The default (1s) would batch ~50 frames per page and the frame-per-page reader would consume the song at ~50× real-time.
• Fast-probe flags. -analyzeduration 0 -probesize 64k on local files cuts ~1-2s of startup latency (default is 5s + 5MB).

A/V sync

• RTP timestamps + RTCP Sender Reports are what synchronise audio and video at the receiver. Pion's default interceptors send SR automatically; the receiver maps RTP timestamps to NTP via the SR.
• Both streamers are started in the same startLocked call (instances/group_call.go), so their wall-clock baselines are within microseconds of each other.
• Each streamer paces by sample.Duration (audio = 20 ms per Opus frame, video = 1/fps per VP8 frame) using a single hoisted timer. Real-time accuracy is sub-millisecond; no drift accumulates over time.
• Don't apply different time-distortion filters to the audio and video legs of one source — e.g., atempo=1.25 on audio without setpts=PTS/1.25 on video. The two will desync linearly at 25 % per minute.
• For RTMP mode (single ffmpeg push) sync is ffmpeg's responsibility — typically not a concern for properly-muxed source files.

Pitfalls

• Requesting video on an audio-only source fails the call. The library opens two ffmpeg subprocesses for Tracks: TrackVideo. If the source has no video stream, ffmpeg's -map 0:v? makes the video leg exit cleanly (no stream), the OGG/IVF parser sees EOF, and the video track is silently skipped — but if audio also fails, you get ErrFile. Don't request video tracks unless you know the container has them.
• Don't switch ffmpeg output back to PCM. It will "work" but defeats the design — you'd be re-encoding in Go (which would require cgo, the very thing this library exists to avoid).
• Raw PCM/YUV is rejected at construction time. FromShell validates the codec/container args and returns ErrInvalidParams with a useful hint pointing at libopus/libvpx.
• SetSource gates on ICE Connected (15 s timeout). ffmpeg is spawned and OGG/IVF headers are parsed outside the lock, but WriteSample only begins after ICE+DTLS completes — samples written before SRTP binding are silently dropped by pion. If ICE fails or the call is stopped during the wait, SetSource returns ErrNotConnected / ErrConnectionFailed.
• Pause in RTMP mode is destructive to the connection. ffmpeg is killed; Telegram drops the RTMP ingest. Resume re-establishes from elapsed_ms. Listeners will see a brief silence.

Performance vs ntgcalls

Both ship into the same Telegram SFU and use the same Opus + VP8 codecs at the same bitrates, so wire bandwidth is identical. The differences are operational, not protocol-level.

Dimension	ntgcalls (libwebrtc, C++)	gotgcall (pion + ffmpeg subprocess, pure Go)
Per-call CPU (steady state, audio-only)	~1–2 % of one core (in-process Opus encoder, no IPC)	~2–4 % of one core (ffmpeg encodes Opus, pipe IPC to Go, pion packetises)
Per-call memory baseline	~15–25 MB (libwebrtc allocators + jitter buffers)	~5–10 MB Go heap + ~10–20 MB per ffmpeg subprocess (drops to ~5 MB on libopus-only audio)
Cold-start to first packet	~50–150 ms (compiled-in encoder ready immediately)	~80–300 ms (ffmpeg spawn + first OGG page; the -analyzeduration 0 -probesize 64k fast-probe flags shave ~1–2 s vs ffmpeg defaults)
Cross-compile / deploy	Requires libwebrtc + glibc + a C++ toolchain on the target ABI; cgo enabled	CGO_ENABLED=0 GOOS=linux GOARCH=arm64 go build — single static binary, scp it, run. ffmpeg as a runtime dep (typically already installed).
Binary size	~20–30 MB (libwebrtc + cgo glue)	~12–18 MB Go binary (no ffmpeg bundled)
Subprocess footprint	None (everything in-process)	1 ffmpeg per call for audio, 2 if video is on (one per leg). Easy to inspect/kill with standard Unix tools; isolates encoder crashes from the bot process.
Pause/resume latency	Mute internal pipeline, sub-ms	Pause: gate the streamer (sub-ms, ffmpeg keeps running). Resume: wake the gate (sub-ms). RTMP mode: kill+restart with -ss, ~100–300 ms gap.
Concurrent calls per process	Bounded by libwebrtc thread pool sizing (~hundreds without tuning)	Bounded by ffmpeg subprocess count + FDs. With WithSharedUDPMux and raised FD limits: tens of thousands. See UDP mux & scaling.
Hot-reload of encoder logic	Recompile + redeploy the whole bot	Swap an ffmpeg flag string at runtime (FromShell) — no rebuild

Trade-offs at a glance:

• ntgcalls wins on raw CPU/memory per call (in-process encoder, no IPC overhead).
• gotgcall wins on operability (static binaries, no C++ chain, ffmpeg-flag flexibility, OS-level subprocess isolation).
• For a typical music bot (10–500 concurrent voice calls), the per-call overhead difference is invisible on any reasonable server.
• For 10 000+ concurrent calls on one box, ntgcalls' lower per-call memory footprint matters; gotgcall offsets some of this by sharing one UDP socket via WithSharedUDPMux.

The actual numbers above are order-of-magnitude estimates; benchmark on your workload before committing to either.

Why pure Go

ntgcalls works fine but pulls in libwebrtc + glibc + a C++ build chain. Cross-compiling music bots becomes a maintenance burden. gotgcall builds with CGO_ENABLED=0 to a single static binary on every supported platform. The trade-off is ffmpeg as a runtime dependency, which most bot deployments already have anyway.

FAQ

Is this a port of ntgcalls / pytgcalls to Go?

No — it's an independent implementation with a deliberately ntgcalls-shaped API so existing bot code translates almost line-for-line. ntgcalls wraps libwebrtc (C++); gotgcall uses pion, the pure-Go WebRTC stack.

Does it work with gogram, MTProto-Go, or other MTProto libraries?

Yes — any of them. The library is blob-only: it produces and consumes JSON strings; you handle the MTProto layer (phone.JoinGroupCall / phone.LeaveGroupCall) in your bot using whichever MTProto Go library you prefer. The examples/bot/ directory has a runnable skeleton against gogram.

Can I use this for a Telegram music bot?

That's the primary use case. See examples/bot/ and the FromShell recipes for piping yt-dlp / atempo / loudness-normalised ffmpeg pipelines.

Does it support video chats / livestreams / RTMP push?

Yes — three modes:

1. WebRTC: send-only audio + video into a normal voice/video chat.
2. RTMP push: "go live" broadcasts to a channel via Telegram's RTMP ingest URL. See RTMP mode.
3. Custom ffmpeg: FromShell / FromShells lets you point at any decodable container or live source — HLS, RTSP, MJPEG, screen capture, etc.

Does it support TGCalls / MTProto E2E voice calls?

No — only group calls and channel RTMP livestreams. 1-on-1 MTProto voice/video calls (TGCalls) require a different signalling path that this library does not currently target.

What Go version is required?

Go 1.26 or newer (uses errors.AsType[T] and a few stdlib refinements added in 1.26).

Does it run on Windows?

Yes. Pure-Go means no Make/gcc/clang. Pause/Resume in WebRTC mode uses a channel gate (works on every OS); RTMP mode uses kill+restart-with--ss (also OS-agnostic — SIGSTOP would be killed by Telegram's RTMP ingest timeout anyway).

How many concurrent calls can one process handle?

The library has no hardcoded limit. The practical ceiling is ffmpeg subprocess count + ICE socket count. Use WithSharedUDPMux() to collapse all calls onto one UDP socket once you're above ~100 concurrent calls.

Where do I report bugs?

Open an issue with logs from WithLogger(slog.New(slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{Level: slog.LevelDebug}))) — debug-level logging covers streamer state, ffmpeg exit, ICE transitions.

See also

• pion/webrtc — pure-Go WebRTC stack (the WebRTC layer underneath).
• amarnathcjd/gogram — pure-Go MTProto client used in the example.
• pytgcalls/ntgcalls — the C++ library this is an alternative to.

License

MIT — see LICENSE.

Documentation

Overview

Package gotgcall is a pure-Go library for streaming audio and video into Telegram group calls. The public API mirrors ntgcalls method names so bot code translates one-to-one.

The library is blob-only: signaling JSON is exchanged through your own MTProto client (typically gogram). Two calls are required:

params, _ := client.CreateCall(chatID)
resp, _   := tg.PhoneJoinGroupCall(... Params: &DataJson{Data: params})
client.Connect(chatID, resp.Updates[...].Call.Params.Data)
client.SetStreamSources(chatID, gotgcall.FromFile("song.mp3", gotgcall.EncodeOptions{}))

See README.md for the full pattern.

Index

• Constants
• Variables
• type CallInfo
• type Client
  • func New(opts ...Option) (*Client, error)
  • func (c *Client) AudioSSRC(chatID int64) (uint32, error)
  • func (c *Client) Calls() map[int64]CallInfo
  • func (c *Client) Close() error
  • func (c *Client) Connect(chatID int64, telegramParams string) error
  • func (c *Client) CreateCall(chatID int64) (string, error)
  • func (c *Client) GetState(chatID int64) (MediaState, error)
  • func (c *Client) Mute(chatID int64) (bool, error)
  • func (c *Client) OnConnectionChange(fn func(chatID int64, info NetworkInfo))
  • func (c *Client) OnMediaStateChange(fn func(chatID int64, state MediaState))
  • func (c *Client) OnStreamEnd(fn func(chatID int64, t StreamType, d Device, err error))
  • func (c *Client) Pause(chatID int64) (bool, error)
  • func (c *Client) Resume(chatID int64) (bool, error)
  • func (c *Client) SetStreamSources(chatID int64, src Source) error
  • func (c *Client) StartRTMP(chatID int64, rtmpURL string) error
  • func (c *Client) Stop(chatID int64) error
  • func (c *Client) Time(chatID int64) (uint64, error)
  • func (c *Client) Unmute(chatID int64) (bool, error)
• type ConnState
• type Device
• type EncodeOptions
• type ICEServer
• type MediaState
• type NetworkInfo
• type NetworkType
• type Option
  • func WithDTLSCertPool(n int) Option
  • func WithDebugLogs() Option
  • func WithDispatchBuffer(n int) Option
  • func WithFFmpegPath(p string) Option
  • func WithFFmpegStderrLog() Option
  • func WithICECandidateLogs() Option
  • func WithICEServers(servers []ICEServer) Option
  • func WithICETimeouts(disconnect, failed, keepalive time.Duration) Option
  • func WithLogger(l *slog.Logger) Option
  • func WithNetworkTypes(types ...NetworkType) Option
  • func WithPionTraceLogs() Option
  • func WithSharedUDPMux() Option
• type SeekableSource
• type Source
• type StreamType
• type Track

Constants

const (
	NetworkTypeUDP4 = webrtc.NetworkTypeUDP4
	NetworkTypeUDP6 = webrtc.NetworkTypeUDP6
	NetworkTypeTCP4 = webrtc.NetworkTypeTCP4
	NetworkTypeTCP6 = webrtc.NetworkTypeTCP6
)

const (
	TrackAudio = media.TrackAudio
	TrackVideo = media.TrackVideo

	Audio      = models.Audio
	Video      = models.Video
	Microphone = models.Microphone
	Camera     = models.Camera

	Connecting   = models.Connecting
	Connected    = models.Connected
	Disconnected = models.Disconnected
	Failed       = models.Failed
	Closed       = models.Closed
)

Variables

var (
	FromFile   = media.FromFile
	FromURL    = media.FromURL
	FromShell  = media.FromShell
	FromShells = media.FromShells
)

var (
	ErrConnectionExists   = models.ErrConnectionExists
	ErrConnectionNotFound = models.ErrConnectionNotFound
	ErrConnectionFailed   = models.ErrConnectionFailed
	ErrInvalidParams      = models.ErrInvalidParams
	ErrFFmpegSpawn        = models.ErrFFmpegSpawn
	ErrFFmpegCrashed      = models.ErrFFmpegCrashed
	ErrFile               = models.ErrFile
	ErrClosed             = models.ErrClosed
	ErrInternal           = models.ErrInternal
	ErrNotConnected       = models.ErrNotConnected
	ErrWrongMode          = models.ErrWrongMode
)

Types

type CallInfo

type CallInfo = models.CallInfo

type Client

type Client struct {
	// contains filtered or unexported fields
}

Client multiplexes many concurrent group calls behind a single process-wide handle. Safe for concurrent use.

func New

func New(opts ...Option) (*Client, error)

New constructs a Client with the given options. Fails fast if the ffmpeg binary isn't on PATH (or wherever WithFFmpegPath points) so callers see the error at startup rather than on first stream.

func (*Client) AudioSSRC

func (c *Client) AudioSSRC(chatID int64) (uint32, error)

AudioSSRC returns the audio SSRC of a WebRTC call. Pass as Source to phone.LeaveGroupCall. Returns ErrWrongMode for RTMP calls.

func (*Client) Calls

func (c *Client) Calls() map[int64]CallInfo

Calls returns a snapshot of all active calls.

func (*Client) Close

func (c *Client) Close() error

Close stops every call and releases resources. Idempotent.

func (*Client) Connect

func (c *Client) Connect(chatID int64, telegramParams string) error

Connect finishes the WebRTC handshake using Telegram's response JSON.

func (*Client) CreateCall

func (c *Client) CreateCall(chatID int64) (string, error)

CreateCall starts a new WebRTC group-call instance for chatID and returns the JSON params the caller must pass to phone.JoinGroupCall.

Concurrent CreateCall / StartRTMP calls for the same chat are serialized; the first one wins, others get ErrConnectionExists without allocating a pion PeerConnection.

func (*Client) GetState

func (c *Client) GetState(chatID int64) (MediaState, error)

GetState returns the current media-state (mute/pause flags).

func (*Client) Mute

func (c *Client) Mute(chatID int64) (bool, error)

func (*Client) OnConnectionChange

func (c *Client) OnConnectionChange(fn func(chatID int64, info NetworkInfo))

OnConnectionChange registers a callback for ICE/DTLS state transitions.

func (*Client) OnMediaStateChange

func (c *Client) OnMediaStateChange(fn func(chatID int64, state MediaState))

OnMediaStateChange registers a callback fired whenever the call's outgoing media state (muted / paused / video-stopped) transitions.

Use this to keep the Telegram-side participant flags (settable via phone.editGroupCallParticipant in your MTProto layer) in sync with the library-side streamer state. Most importantly: when /play (audio-only) is followed by /vplay (video) on the SAME call, this fires with state.VideoStopped=false, signaling that you should flip the participant's video_stopped flag MTProto-side. Without that signal, Telegram's SFU may drop the late video even though our RTP is correct.

Mirror of ntgcalls' onUpgrade(MediaState) pattern. Fires on the dispatcher goroutine, safe to re-enter the Client API from within.

func (*Client) OnStreamEnd

func (c *Client) OnStreamEnd(fn func(chatID int64, t StreamType, d Device, err error))

OnStreamEnd registers a callback fired when a track ends (EOF, crash, stop). Called on the dispatcher goroutine so it is safe to re-enter the Client API from within.

func (*Client) Pause

func (c *Client) Pause(chatID int64) (bool, error)

func (*Client) Resume

func (c *Client) Resume(chatID int64) (bool, error)

func (*Client) SetStreamSources

func (c *Client) SetStreamSources(chatID int64, src Source) error

SetStreamSources installs or replaces the streaming source for chatID. Encode options (FPS, tracks, bitrates) ride along with the Source — set them on the constructor (FromFile/FromURL).

func (*Client) StartRTMP

func (c *Client) StartRTMP(chatID int64, rtmpURL string) error

StartRTMP creates an RTMP-push call for chatID. The caller obtains rtmpURL via phone.GetGroupCallStreamRtmpUrl gogram-side. Serialised with CreateCall via the same per-chat creation mutex.

func (*Client) Stop

func (c *Client) Stop(chatID int64) error

Stop tears down the call and clears every per-chat scrap of state the library kept (call instance, create-mutex). After Stop the chatID can be re-used cleanly.

func (*Client) Time

func (c *Client) Time(chatID int64) (uint64, error)

Time returns elapsed ms of media pushed.

func (*Client) Unmute

func (c *Client) Unmute(chatID int64) (bool, error)

type ConnState

type ConnState = models.ConnState

type Device

type Device = models.Device

type EncodeOptions

type EncodeOptions = media.EncodeOptions

type ICEServer

type ICEServer = webrtc.ICEServer

ICEServer is re-exported so callers can configure STUN/TURN without importing pion directly.

type MediaState

type MediaState = models.MediaState

type NetworkInfo

type NetworkInfo = models.NetworkInfo

type NetworkType

type NetworkType = webrtc.NetworkType

NetworkType is re-exported for WithNetworkTypes.

type Option

type Option func(*config)

func WithDTLSCertPool

func WithDTLSCertPool(n int) Option

WithDTLSCertPool sets the size of the pre-generated DTLS certificate pool. Larger pools absorb bigger call-creation bursts without keygen latency. 0 disables pre-generation.

func WithDebugLogs

func WithDebugLogs() Option

WithDebugLogs is a convenience that installs a Debug-level text handler writing to os.Stderr. Equivalent to:

WithLogger(slog.New(slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{Level: slog.LevelDebug})))

Use this when reporting bugs — debug-level output covers ICE/DTLS state, ffmpeg exit codes, streamer pacing, and pion-internal events bridged through the new pion→slog adapter.

func WithDispatchBuffer

func WithDispatchBuffer(n int) Option

WithDispatchBuffer sizes the event dispatcher's channel. Default 256.

func WithFFmpegPath

func WithFFmpegPath(p string) Option

WithFFmpegPath overrides the ffmpeg binary path (default "ffmpeg").

func WithFFmpegStderrLog

func WithFFmpegStderrLog() Option

WithFFmpegStderrLog tees ffmpeg's stderr output to the library logger at Debug level while the process is running. Without this, ffmpeg stderr is only surfaced in the final error message (last 512 bytes) when the subprocess crashes — useful for crash diagnosis but useless for "ffmpeg is running but I see no audio" symptoms. Enable for verbose diagnosis.

func WithICECandidateLogs

func WithICECandidateLogs() Option

WithICECandidateLogs logs every locally-gathered ICE candidate (host / srflx / relay, address, port, foundation) at Debug level via the PeerConnection's OnICECandidate hook. Pairs well with WithPionTraceLogs for "why is ICE failing" diagnosis: this option shows what we offered, pion-trace shows which pairs were tried, and the remote answer's candidate list (parsed in jsonparams) shows what Telegram returned.

func WithICEServers

func WithICEServers(servers []ICEServer) Option

WithICEServers overrides the default ICE server list (2 Google STUN servers). Pass TURN entries for users behind symmetric NAT or restrictive firewalls. Pass an empty slice to disable STUN entirely (host-only candidates).

gotgcall.WithICEServers([]gotgcall.ICEServer{
    {URLs: []string{"turn:turn.example.com:3478"},
     Username: "u", Credential: "p"},
})

func WithICETimeouts

func WithICETimeouts(disconnect, failed, keepalive time.Duration) Option

WithICETimeouts overrides pion's ICE timing. Pass 0 for any value to keep the library default (60s disconnect grace / 120s failed / 2s keepalive). Telegram's edge wobble on rejoin often takes 60-90s to settle, hence the generous defaults. Use longer values on unstable networks where brief connectivity drops shouldn't kill the call; pass shorter values for ultra-responsive UIs that need faster fail-detection.

func WithLogger

func WithLogger(l *slog.Logger) Option

WithLogger sets a structured logger for internal events.

func WithNetworkTypes

func WithNetworkTypes(types ...NetworkType) Option

WithNetworkTypes overrides the ICE candidate network-type whitelist. Default is UDP4+UDP6 (matching ntgcalls' PORTALLOCATOR_ENABLE_IPV6). Telegram's SFU accepts IPv6 candidates and dual-stack hosts get more candidate pairs. Add TCP for restrictive environments where UDP is blocked.

gotgcall.WithNetworkTypes(
    gotgcall.NetworkTypeUDP4,
    gotgcall.NetworkTypeUDP6,
    gotgcall.NetworkTypeTCP4,
)

func WithPionTraceLogs

func WithPionTraceLogs() Option

WithPionTraceLogs remaps pion's Trace-level output (per-ICE-check, per- candidate-pair, per-binding-request) to slog.LevelDebug instead of the default sub-debug level. Use this when ICE is stuck in "Checking" and you need to see exactly which candidate pairs are being tried, which fail, and which (if any) get a response from the remote.

gotgcall.New(gotgcall.WithDebugLogs(), gotgcall.WithPionTraceLogs())

Volume warning: ICE Trace at scale is several hundred lines per call. Use for diagnosis, not steady-state production.

func WithSharedUDPMux

func WithSharedUDPMux() Option

WithSharedUDPMux makes all calls share one UDP socket for ICE traffic. Useful for high-concurrency setups (100+ simultaneous calls).

type SeekableSource

type SeekableSource = media.SeekableSource

type Source

type Source = media.Source

type StreamType

type StreamType = models.StreamType

type Track

type Track = media.Track