diff --git a/daemon/audio_player.c b/daemon/audio_player.c
index f3f79926e..ab5c829ba 100644
--- a/daemon/audio_player.c
+++ b/daemon/audio_player.c
@@ -33,6 +33,12 @@ static bool audio_player_run(struct media_player *mp) {
 	unsigned int size;
 	void *buf = mix_buffer_read_fast(&ap->mb, ap->ptime, &size);
 	if (!buf) {
+		if (!size) {
+			// error or not active: just reschedule
+			timeval_add_usec(&mp->next_run, ap->ptime_us);
+			timerthread_obj_schedule_abs(&mp->tt_obj, &mp->next_run);
+			return false;
+		}
 		buf = g_alloca(size);
 		mix_buffer_read_slow(&ap->mb, buf, ap->ptime);
 	}
@@ -64,7 +70,6 @@ bool audio_player_setup(struct call_media *m, const struct rtp_payload_type *dst
 	unsigned int ptime_smp = ptime_ms * clockrate / 1000; // in samples
 
 	// TODO: shortcut this to avoid the detour of avframe -> avpacket -> avframe (all in s16)
-	// TODO: determine dest sample format from created encoder
 	struct rtp_payload_type src_pt = {
 		.payload_type = -1,
 		.encoding = STR_CONST_INIT("PCM-S16LE"), // XXX support flp
@@ -121,7 +126,8 @@ bool audio_player_setup(struct call_media *m, const struct rtp_payload_type *dst
 
 	bufsize_ms = MAX(bufsize_ms, ptime_ms * 2); // make sure the buf size is at least 2 frames
 
-	mix_buffer_init(&ap->mb, AV_SAMPLE_FMT_S16, clockrate, dst_pt->channels, bufsize_ms, delay_ms);
+	mix_buffer_init_active(&ap->mb, AV_SAMPLE_FMT_S16, clockrate, dst_pt->channels, bufsize_ms, delay_ms,
+			false);
 
 	return true;
 
@@ -131,6 +137,16 @@ error:
 }
 
 
+void audio_player_activate(struct call_media *m) {
+	if (!m)
+		return;
+	struct audio_player *ap = m->audio_player;
+	if (!ap)
+		return;
+	mix_buffer_activate(&ap->mb);
+}
+
+
 // call locked in W
 void audio_player_start(struct call_media *m) {
 	struct audio_player *ap;
diff --git a/daemon/call.c b/daemon/call.c
index 4919a9491..8e30eca1e 100644
--- a/daemon/call.c
+++ b/daemon/call.c
@@ -2759,6 +2759,9 @@ void codecs_offer_answer(struct call_media *media, struct call_media *other_medi
 			set_transcoding_flag(media->monologue, other_media->monologue, true);
 		if (codec_handlers_update(other_media, media, NULL, NULL))
 			set_transcoding_flag(other_media->monologue, media->monologue, true);
+
+		// activate audio player if needed (not done by codec_handlers_update without `flags`)
+		audio_player_activate(media);
 	}
 }
 
diff --git a/daemon/call_interfaces.c b/daemon/call_interfaces.c
index ce7d4ff11..232254ee5 100644
--- a/daemon/call_interfaces.c
+++ b/daemon/call_interfaces.c
@@ -975,6 +975,9 @@ static void call_ng_flags_flags(struct sdp_ng_flags *out, str *s, void *dummy) {
 		case CSH_LOOKUP("reset"):
 			out->reset = 1;
 			break;
+		case CSH_LOOKUP("early-media"):
+			out->early_media = 1;
+			break;
 		case CSH_LOOKUP("all"):
 			out->all = ALL_ALL;
 			break;
diff --git a/daemon/codec.c b/daemon/codec.c
index f040c4785..574ea267b 100644
--- a/daemon/codec.c
+++ b/daemon/codec.c
@@ -1438,6 +1438,8 @@ next:
 
 			audio_player_setup(sink, pref_dest_codec, rtpe_config.audio_buffer_length,
 					rtpe_config.audio_buffer_delay);
+			if (flags && (flags->early_media || flags->opmode == OP_ANSWER))
+				audio_player_activate(sink);
 		}
 	}
 
diff --git a/docs/ng_control_protocol.md b/docs/ng_control_protocol.md
index 712327784..8f33b5217 100644
--- a/docs/ng_control_protocol.md
+++ b/docs/ng_control_protocol.md
@@ -805,6 +805,14 @@ Spaces in each string may be replaced by hyphens.
 	the DSP to detect in-band DTMF audio tones even when it
 	wouldn't otherwise be necessary.
 
+* `early media`
+
+    Used in conjunction with the audio player. If set, audio playback is
+    started immediately when processing an `offer` message. The default
+    behaviour is to start the audio player only after the `answer` has been
+    processed, or when any audio to be played back has actually been received
+    (either from another party to the call, or via the `play media` command).
+
 * `full rtcp attribute`
 
 	Include the full version of the `a=rtcp` line (complete with network address) instead of
diff --git a/include/audio_player.h b/include/audio_player.h
index af8bdfc08..7bf2176ec 100644
--- a/include/audio_player.h
+++ b/include/audio_player.h
@@ -22,6 +22,7 @@ struct rtp_payload_type;
 
 bool audio_player_setup(struct call_media *, const struct rtp_payload_type *,
 		unsigned int size_ms, unsigned int delay_ms);
+void audio_player_activate(struct call_media *);
 void audio_player_free(struct call_media *);
 
 void audio_player_start(struct call_media *);
@@ -36,6 +37,7 @@ void audio_player_add_frame(struct audio_player *, uint32_t ssrc, AVFrame *);
 INLINE void audio_player_start(struct call_media *m) { }
 INLINE void audio_player_free(struct call_media *m) { }
 INLINE void audio_player_stop(struct call_media *m) { }
+INLINE void audio_player_activate(struct call_media *m) { }
 
 #endif
 
diff --git a/include/call_interfaces.h b/include/call_interfaces.h
index 083bf8186..0e9beeb4e 100644
--- a/include/call_interfaces.h
+++ b/include/call_interfaces.h
@@ -162,6 +162,7 @@ struct sdp_ng_flags {
 	             single_codec:1,
 		     reuse_codec:1,
 		     allow_transcoding:1,
+		     early_media:1,
 		     accept_any:1,
 	             inject_dtmf:1,
 		     detect_dtmf:1,
diff --git a/lib/mix_buffer.c b/lib/mix_buffer.c
index 51778e92e..cbb3d13e8 100644
--- a/lib/mix_buffer.c
+++ b/lib/mix_buffer.c
@@ -73,8 +73,8 @@ static void fill_up_to(struct mix_buffer *mb, unsigned int up_to) {
 void *mix_buffer_read_fast(struct mix_buffer *mb, unsigned int samples, unsigned int *size) {
 	LOCK(&mb->lock);
 
-	if (samples > mb->size) {
-		*size = 0; // error
+	if (samples > mb->size || !mb->active) {
+		*size = 0; // error or inactive
 		return NULL;
 	}
 
@@ -245,6 +245,8 @@ bool mix_buffer_write_delay(struct mix_buffer *mb, uint32_t ssrc, const void *bu
 	if (created)
 		mix_buff_src_shift_delay(mb, src, last, now);
 
+	mb->active = true;
+
 	// loop twice at the most to re-run logic after a reset
 	while (true) {
 		// shortcut if we're at the write head
@@ -284,8 +286,8 @@ static struct ssrc_entry *mix_buffer_ssrc_new(void *p) {
 
 
 // struct must be zeroed already
-bool mix_buffer_init(struct mix_buffer *mb, enum AVSampleFormat fmt, unsigned int clockrate,
-		unsigned int channels, unsigned int size_ms, unsigned int delay_ms)
+bool mix_buffer_init_active(struct mix_buffer *mb, enum AVSampleFormat fmt, unsigned int clockrate,
+		unsigned int channels, unsigned int size_ms, unsigned int delay_ms, bool active)
 {
 	switch (fmt) {
 		case AV_SAMPLE_FMT_S16:
@@ -305,6 +307,7 @@ bool mix_buffer_init(struct mix_buffer *mb, enum AVSampleFormat fmt, unsigned in
 	mb->clockrate = clockrate;
 	mb->channels = channels;
 	mb->delay = delay;
+	mb->active = active;
 
 	mb->ssrc_hash = create_ssrc_hash_full_fast(mix_buffer_ssrc_new, mb);
 
diff --git a/lib/mix_buffer.h b/lib/mix_buffer.h
index 8b422ab4a..9328d5732 100644
--- a/lib/mix_buffer.h
+++ b/lib/mix_buffer.h
@@ -50,6 +50,7 @@ struct mix_buffer {
 	unsigned int delay; // initial write delay for new inputs/sources
 
 	unsigned int loops; // how many times the write pos has circled around
+	bool active; // to optionally suppress early media
 
 	// implementation details
 	const struct mix_buffer_impl *impl;
@@ -58,8 +59,14 @@ struct mix_buffer {
 };
 
 
-bool mix_buffer_init(struct mix_buffer *, enum AVSampleFormat, unsigned int clockrate,
-		unsigned int channels, unsigned int size_ms, unsigned int delay_ms);
+bool mix_buffer_init_active(struct mix_buffer *, enum AVSampleFormat, unsigned int clockrate,
+		unsigned int channels, unsigned int size_ms, unsigned int delay_ms, bool active);
+#define mix_buffer_init(mb, fmt, clockrate, channels, size_ms, delay_ms) \
+	mix_buffer_init_active(mb, fmt, clockrate, channels, size_ms, delay_ms, true)
+INLINE void mix_buffer_activate(struct mix_buffer *mb) {
+	LOCK(&mb->lock);
+	mb->active = true;
+}
 void mix_buffer_destroy(struct mix_buffer *);
 
 void *mix_buffer_read_fast(struct mix_buffer *, unsigned int samples, unsigned int *size);
diff --git a/utils/rtpengine-ng-client b/utils/rtpengine-ng-client
index 70a249044..691b3bae5 100755
--- a/utils/rtpengine-ng-client
+++ b/utils/rtpengine-ng-client
@@ -49,6 +49,7 @@ my @flags = qw(
 	passthrough
 	no-passthrough
 	pause
+	early-media
 );
 
 my @string_opts = qw(