Datasets:

VoiceHub
/

mytts-en-samples

checkpoint stringclasses 14 values	run stringclasses 14 values	step int64 30k 30k	item stringclasses 5 values	category stringclasses 5 values	voice_set stringclasses 1 value	voice stringclasses 5 values	prompt_file stringclasses 5 values	prompt_text stringclasses 5 values	text stringclasses 5 values	duration_s float64 5.12 17.7	watermark stringclasses 1 value	watermark_score float64 1 1	cfg_w float64 4 4	steps int64 32 32	noise_scale float64 0.9 0.9	echo_dev_wer float64 0.01 0.03	echo_dev_sim_o float64 0.61 0.69	echo_dev_utmos float64 2.55 3.09	seed_dev_wer float64 0.05 0.08	seed_dev_sim_o float64 0.37 0.45	seed_dev_utmos float64 2.34 2.99	en_e1_verdict stringclasses 6 values
en55-t06/step_0030000	en55-t06	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.033379	0.675899	3.086876	0.08013	0.448403	2.960885	worse on echo_dev, seed_dev
en55-t06/step_0030000	en55-t06	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.033379	0.675899	3.086876	0.08013	0.448403	2.960885	worse on echo_dev, seed_dev
en55-t06/step_0030000	en55-t06	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	0.9993	4	32	0.9	0.033379	0.675899	3.086876	0.08013	0.448403	2.960885	worse on echo_dev, seed_dev
en55-t06/step_0030000	en55-t06	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.033379	0.675899	3.086876	0.08013	0.448403	2.960885	worse on echo_dev, seed_dev
en55-t06/step_0030000	en55-t06	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.033379	0.675899	3.086876	0.08013	0.448403	2.960885	worse on echo_dev, seed_dev
en55-pc06-t06/step_0030000	en55-pc06-t06	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.021409	0.687277	3.085622	0.065592	0.448988	2.985143	worse on seed_dev
en55-pc06-t06/step_0030000	en55-pc06-t06	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.021409	0.687277	3.085622	0.065592	0.448988	2.985143	worse on seed_dev
en55-pc06-t06/step_0030000	en55-pc06-t06	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.021409	0.687277	3.085622	0.065592	0.448988	2.985143	worse on seed_dev
en55-pc06-t06/step_0030000	en55-pc06-t06	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.021409	0.687277	3.085622	0.065592	0.448988	2.985143	worse on seed_dev
en55-pc06-t06/step_0030000	en55-pc06-t06	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.021409	0.687277	3.085622	0.065592	0.448988	2.985143	worse on seed_dev
en55-pc06-t08/step_0030000	en55-pc06-t08	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.681454	3.039311	0.052975	0.439728	2.938725	no significant gain
en55-pc06-t08/step_0030000	en55-pc06-t08	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.681454	3.039311	0.052975	0.439728	2.938725	no significant gain
en55-pc06-t08/step_0030000	en55-pc06-t08	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.681454	3.039311	0.052975	0.439728	2.938725	no significant gain
en55-pc06-t08/step_0030000	en55-pc06-t08	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.681454	3.039311	0.052975	0.439728	2.938725	no significant gain
en55-pc06-t08/step_0030000	en55-pc06-t08	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.681454	3.039311	0.052975	0.439728	2.938725	no significant gain
en55-pc06-t10/step_0030000	en55-pc06-t10	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.020718	0.662944	2.953989	0.052139	0.413289	2.788466	no significant gain
en55-pc06-t10/step_0030000	en55-pc06-t10	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.020718	0.662944	2.953989	0.052139	0.413289	2.788466	no significant gain
en55-pc06-t10/step_0030000	en55-pc06-t10	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.020718	0.662944	2.953989	0.052139	0.413289	2.788466	no significant gain
en55-pc06-t10/step_0030000	en55-pc06-t10	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.020718	0.662944	2.953989	0.052139	0.413289	2.788466	no significant gain
en55-pc06-t10/step_0030000	en55-pc06-t10	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.020718	0.662944	2.953989	0.052139	0.413289	2.788466	no significant gain
en55-t08/step_0030000	en55-t08	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.03384	0.662396	2.925687	0.08013	0.435471	2.777853	worse on echo_dev, seed_dev
en55-t08/step_0030000	en55-t08	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.03384	0.662396	2.925687	0.08013	0.435471	2.777853	worse on echo_dev, seed_dev
en55-t08/step_0030000	en55-t08	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.03384	0.662396	2.925687	0.08013	0.435471	2.777853	worse on echo_dev, seed_dev
en55-t08/step_0030000	en55-t08	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.03384	0.662396	2.925687	0.08013	0.435471	2.777853	worse on echo_dev, seed_dev
en55-t08/step_0030000	en55-t08	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.03384	0.662396	2.925687	0.08013	0.435471	2.777853	worse on echo_dev, seed_dev
en29-fixed-pc06-tf04/step_0030000	en29-fixed-pc06-tf04	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.012431	0.648668	2.758563	0.059659	0.406573	2.582931	adopted: adopt
en29-fixed-pc06-tf04/step_0030000	en29-fixed-pc06-tf04	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.012431	0.648668	2.758563	0.059659	0.406573	2.582931	adopted: adopt
en29-fixed-pc06-tf04/step_0030000	en29-fixed-pc06-tf04	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.012431	0.648668	2.758563	0.059659	0.406573	2.582931	adopted: adopt
en29-fixed-pc06-tf04/step_0030000	en29-fixed-pc06-tf04	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.012431	0.648668	2.758563	0.059659	0.406573	2.582931	adopted: adopt
en29-fixed-pc06-tf04/step_0030000	en29-fixed-pc06-tf04	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.012431	0.648668	2.758563	0.059659	0.406573	2.582931	adopted: adopt
en29-fixed/step_0030000	en29-fixed	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.017265	0.639268	2.757469	0.052139	0.392305	2.654783	no significant gain
en29-fixed/step_0030000	en29-fixed	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.017265	0.639268	2.757469	0.052139	0.392305	2.654783	no significant gain
en29-fixed/step_0030000	en29-fixed	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	0.9995	4	32	0.9	0.017265	0.639268	2.757469	0.052139	0.392305	2.654783	no significant gain
en29-fixed/step_0030000	en29-fixed	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.017265	0.639268	2.757469	0.052139	0.392305	2.654783	no significant gain
en29-fixed/step_0030000	en29-fixed	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.017265	0.639268	2.757469	0.052139	0.392305	2.654783	no significant gain
en29-fixed-pc06/step_0030000	en29-fixed-pc06	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.010589	0.651559	2.710976	0.063001	0.423566	2.630353	adopted: adopt
en29-fixed-pc06/step_0030000	en29-fixed-pc06	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.010589	0.651559	2.710976	0.063001	0.423566	2.630353	adopted: adopt
en29-fixed-pc06/step_0030000	en29-fixed-pc06	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.010589	0.651559	2.710976	0.063001	0.423566	2.630353	adopted: adopt
en29-fixed-pc06/step_0030000	en29-fixed-pc06	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.010589	0.651559	2.710976	0.063001	0.423566	2.630353	adopted: adopt
en29-fixed-pc06/step_0030000	en29-fixed-pc06	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.010589	0.651559	2.710976	0.063001	0.423566	2.630353	adopted: adopt
en55-repafade/step_0030000	en55-repafade	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.031538	0.628023	2.702013	0.072777	0.398135	2.513215	worse on echo_dev, seed_dev
en55-repafade/step_0030000	en55-repafade	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.031538	0.628023	2.702013	0.072777	0.398135	2.513215	worse on echo_dev, seed_dev
en55-repafade/step_0030000	en55-repafade	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.031538	0.628023	2.702013	0.072777	0.398135	2.513215	worse on echo_dev, seed_dev
en55-repafade/step_0030000	en55-repafade	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.031538	0.628023	2.702013	0.072777	0.398135	2.513215	worse on echo_dev, seed_dev
en55-repafade/step_0030000	en55-repafade	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.031538	0.628023	2.702013	0.072777	0.398135	2.513215	worse on echo_dev, seed_dev
en-t1-base-s2/step_0030000	en-t1-base-s2	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.025092	0.629718	2.696644	0.058406	0.391778	2.623275	baseline (seed 2); vs the other seed: no significant gain
en-t1-base-s2/step_0030000	en-t1-base-s2	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.025092	0.629718	2.696644	0.058406	0.391778	2.623275	baseline (seed 2); vs the other seed: no significant gain
en-t1-base-s2/step_0030000	en-t1-base-s2	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.025092	0.629718	2.696644	0.058406	0.391778	2.623275	baseline (seed 2); vs the other seed: no significant gain
en-t1-base-s2/step_0030000	en-t1-base-s2	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	0.9985	4	32	0.9	0.025092	0.629718	2.696644	0.058406	0.391778	2.623275	baseline (seed 2); vs the other seed: no significant gain
en-t1-base-s2/step_0030000	en-t1-base-s2	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.025092	0.629718	2.696644	0.058406	0.391778	2.623275	baseline (seed 2); vs the other seed: no significant gain
en-t1-base/step_0030000	en-t1-base	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.019567	0.63101	2.677675	0.059576	0.392638	2.583175	baseline (seed 1234)
en-t1-base/step_0030000	en-t1-base	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.019567	0.63101	2.677675	0.059576	0.392638	2.583175	baseline (seed 1234)
en-t1-base/step_0030000	en-t1-base	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.019567	0.63101	2.677675	0.059576	0.392638	2.583175	baseline (seed 1234)
en-t1-base/step_0030000	en-t1-base	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.019567	0.63101	2.677675	0.059576	0.392638	2.583175	baseline (seed 1234)
en-t1-base/step_0030000	en-t1-base	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.019567	0.63101	2.677675	0.059576	0.392638	2.583175	baseline (seed 1234)
en25-hubl24/step_0030000	en25-hubl24	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.617013	2.592628	0.053727	0.381315	2.452491	no significant gain
en25-hubl24/step_0030000	en25-hubl24	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.617013	2.592628	0.053727	0.381315	2.452491	no significant gain
en25-hubl24/step_0030000	en25-hubl24	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.617013	2.592628	0.053727	0.381315	2.452491	no significant gain
en25-hubl24/step_0030000	en25-hubl24	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.617013	2.592628	0.053727	0.381315	2.452491	no significant gain
en25-hubl24/step_0030000	en25-hubl24	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.016805	0.617013	2.592628	0.053727	0.381315	2.452491	no significant gain
en25-hubl18/step_0030000	en25-hubl18	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.020258	0.607467	2.590482	0.054896	0.369111	2.537227	no significant gain
en25-hubl18/step_0030000	en25-hubl18	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.020258	0.607467	2.590482	0.054896	0.369111	2.537227	no significant gain
en25-hubl18/step_0030000	en25-hubl18	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.020258	0.607467	2.590482	0.054896	0.369111	2.537227	no significant gain
en25-hubl18/step_0030000	en25-hubl18	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.020258	0.607467	2.590482	0.054896	0.369111	2.537227	no significant gain
en25-hubl18/step_0030000	en25-hubl18	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.020258	0.607467	2.590482	0.054896	0.369111	2.537227	no significant gain
en25-hubl24fp32/step_0030000	en25-hubl24fp32	30,000	01-short	short	echo	spk_0610	prompts/01-short.wav	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.	5.12	audioseal_wm_16bits	1	4	32	0.9	0.017495	0.61378	2.552274	0.058155	0.39425	2.337512	no significant gain
en25-hubl24fp32/step_0030000	en25-hubl24fp32	30,000	02-question	question	echo	spk_0342	prompts/02-question.wav	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?	7.04	audioseal_wm_16bits	1	4	32	0.9	0.017495	0.61378	2.552274	0.058155	0.39425	2.337512	no significant gain
en25-hubl24fp32/step_0030000	en25-hubl24fp32	30,000	03-numbers	numbers	echo	spk_2669	prompts/03-numbers.wav	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.	8.4	audioseal_wm_16bits	1	4	32	0.9	0.017495	0.61378	2.552274	0.058155	0.39425	2.337512	no significant gain
en25-hubl24fp32/step_0030000	en25-hubl24fp32	30,000	04-conversational	conversational	echo	spk_3757	prompts/04-conversational.wav	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.	11.76	audioseal_wm_16bits	1	4	32	0.9	0.017495	0.61378	2.552274	0.058155	0.39425	2.337512	no significant gain
en25-hubl24fp32/step_0030000	en25-hubl24fp32	30,000	05-long	long	echo	spk_1437	prompts/05-long.wav	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.	17.7	audioseal_wm_16bits	1	4	32	0.9	0.017495	0.61378	2.552274	0.058155	0.39425	2.337512	no significant gain

Archive: screening models (80 M, 88.8 h, 30k steps) used to choose the recipe. The main model, DACFlow-EN-10k: VoiceHub/DACFlow-EN-10k · its training data (tokenized): VoiceHub/DACFlow-EN-10k-data

mytts-en samples

Read this first: every sample on this page comes from a small screening model, not the planned model. The 14 runs published so far are Tier-1 screening runs: ~80 M parameters, 30k training steps (about 1-1.5 GPU-hours each), trained on only 88.8 hours of speech (31,987 clips, 3,587 voices), 0.8 % of the provided ~11,150-hour SynDataLab-EN/echo-clones-4m-en. Screening runs exist only to choose settings (each run changes one technique and is compared with two baseline seeds), so their quality is far below what the main model is planned to reach.

The main model (~206 M parameters, the full ~11,150 h, 200k steps) has its own page: VoiceHub/DACFlow-EN-10k (weights, listening samples and scores; the data: VoiceHub/DACFlow-EN-10k-data).

Listening samples of every checkpoint of VoiceHub/mytts-en, a hobby project: a zero-shot English text-to-speech model (a flow-matching diffusion transformer on Semantic-DACVAE latents, 48 kHz) that clones a voice from a short prompt. Trained only on the provided data: SynDataLab-EN/echo-clones-4m-en, synthetic EchoTTS speech (no external corpus). Only checkpoints trained exclusively on that data are shown.

Every checkpoint reads the same 5 prompts and texts (one per category: a short everyday sentence, a question, numbers / dates / abbreviations, a medium conversational turn, a long two-sentence passage), each prompt spoken by a different held-out voice of the provided data that no checkpoint was trained on, so checkpoints and runs can be compared item by item.

Start here

The checkpoints worth hearing first, side by side (the sections below have every published checkpoint):

en55-pc06-t08 step 30000: the Tier-1 arm of the main-run (M4) recipe, and the best-sounding screening run that is no worse than the baseline on WER (the highest echo-dev UTMOS among the finished runs whose EN-E1 verdict is not "worse"): EN-55 noise-schedule arm with the EN-29 pc06 prompting: t_mean -0.8 / t_std 0.8 (EN-E1: no significant gain). echo-dev WER 1.68 % · SIM-o 0.681 · UTMOS 3.04 (GT 4.20) | seed-dev WER 5.30 % · SIM-o 0.440 · UTMOS 2.94 (GT 3.52)
en-t1-base step 30000: the Tier-1 baseline (seed 1234), the reference every screening arm is judged against. echo-dev WER 1.96 % · SIM-o 0.631 · UTMOS 2.68 (GT 4.20) | seed-dev WER 5.96 % · SIM-o 0.393 · UTMOS 2.58 (GT 3.52)

Item	Category	en55-pc06-t08 step 30000	en-t1-base step 30000
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

How to load

# pip install datasets torchcodec  (datasets >= 4 decodes audio with torchcodec)
from datasets import load_dataset

ds = load_dataset("VoiceHub/mytts-en-samples", split="train")  # downloads metadata.jsonl, samples/ and prompts/
row = ds[0]  # row["audio"] (the sample), row["prompt_audio"], row["text"], row["prompt_text"], row["checkpoint"], row["echo_dev_wer"], ...

Or listen below; the dataset viewer also plays every sample (audio) next to its prompt (prompt_audio).

How the samples are made

Each checkpoint folder of the model repo is loaded with Synthesizer.from_export (the pickle-free release, verified against the training export before upload) and reads every item with the EN-E1 evaluation settings: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule, sampling seed 0 for every item (the run's own CFG when its evaluation used another one; each section says).
No selection or retries: one draw per item, exactly what the checkpoint produced.
Every sample carries the AudioSeal watermark (16-bit message 0100110101010100), checked with the AudioSeal detector before upload (watermark_score in the metadata). The prompts are the original clips of the provided data, unmarked.

Prompts

Item	Category	Voice	Prompt transcript	Text every checkpoint reads
`01-short`	short everyday sentence	spk_0610 · higher pitch (~202 Hz)	I keep telling myself I just need to power through, like, push a little harder, and then it'll be fine. But it's not getting fine. It's getting— it's getting worse.	I left my umbrella at the office again, so I'm definitely getting soaked on the way home.
`02-question`	question	spk_0342 · lower pitch (~149 Hz)	Wait, that mole— has it always looked like that? No, stop.	Have you ever noticed that the quietest person in the room usually has the most interesting story to tell?
`03-numbers`	numbers, dates, abbreviations	spk_2669 · higher pitch (~181 Hz)	this woman on the train was literally eating a whole bag of chips, like, crunching so loud, and I'm sitting there like, okay, do I move? whatever.	Dr. Patel moved my appointment to Tuesday, March 3rd, at 4:15 p.m., and the co-pay went up from $20 to $35.
`04-conversational`	medium conversational turn (~10 s)	spk_3757 · lower pitch (~115 Hz)	I'm sorry, but the subway doors closed on my bag. Again.	So I finally tried that new ramen place downtown, and honestly, it was worth the wait. The broth was amazing, but next time I'm definitely skipping the extra spicy option.
`05-long`	long passage (~20 s, two sentences)	spk_1437 · deepest voice (~108 Hz)	Honestly I think we need to just call IT and have them... you know, actually fix it this time. I'm too old for this.	When the storm passed, the whole neighborhood came outside to look at the damage, and although a few fences had fallen and the old oak tree had lost its biggest branch, everyone was relieved that nobody was hurt. They spent the rest of the afternoon clearing the street and sharing whatever food they had left.

Voices: held-out echo-clean dev voices of SynDataLab-EN/echo-clones-4m-en (pitch = the prompt's median F0), from the dev list configs/eval/en/echo_heldout_dev.txt: never in a training catalog. They are not the voices of the EN-E1 echo-dev set behind the model card's echo-dev numbers (those are held-out voices of echo_heldout_test.txt).

Checkpoints

The main model first (once it trains), then the finished screening runs (the highest echo-dev UTMOS first), then any screening run still training; grouped by run, newest step first. The newest checkpoint of each run is open, its earlier steps are collapsed. UTMOS is shown with the ground truth's (GT) on the same set: the ceiling. Shown: of the main model every checkpoint from step 20k on; of each screening run only its final checkpoint (the one its EN-E1 evaluation scores). Checkpoints before step 20k are not published: they have not learned to align text and speech yet (on the five listening items, a mean WER of 82-111 % and UTMOS 1.3-1.7 at 5k-10k steps, against 5.5 % and 2.8 at 30k). A screening run's later steps before its final one have learned to align too; they are left out only so that each run is heard at the checkpoint it is judged by.

Screening runs:

Run en55-t06: EN-55 noise-schedule arm: t_mean -0.6 / t_std 0.9 (rejected: worse on echo_dev, seed_dev) · 30000
Run en55-pc06-t06: EN-55 noise-schedule arm with the EN-29 pc06 prompting: t_mean -0.6 / t_std 0.9 (rejected: worse on seed_dev) · 30000
Run en55-pc06-t08: EN-55 noise-schedule arm with the EN-29 pc06 prompting: t_mean -0.8 / t_std 0.8 (EN-E1: no significant gain; the Tier-1 arm of the main-run (M4) recipe) · 30000
Run en55-pc06-t10: EN-55 noise-schedule arm with the EN-29 pc06 prompting: t_mean -1.0 / t_std 0.8 (rejected: no significant gain) · 30000
Run en55-t08: EN-55 noise-schedule arm: t_mean -0.8 (rejected: worse on echo_dev, seed_dev) · 30000
Run en29-fixed-pc06-tf04: EN-29 prompting arm: fixed picker, p_cross 0.6, text drop 0.4 (adopted) · 30000
Run en29-fixed: EN-29 prompting arm: fixed cross-prompt picker (rejected: no significant gain) · 30000
Run en29-fixed-pc06: EN-29 prompting arm: fixed picker, p_cross 0.6 (adopted) · 30000
Run en55-repafade: EN-55 arm: REPA faded out after 18k steps (rejected: worse on echo_dev, seed_dev) · 30000
Run en-t1-base-s2: baseline, seed 2 · 30000
Run en-t1-base: baseline, seed 1234 · 30000
Run en25-hubl24: EN-25 REPA-teacher arm: HuBERT-large layer 24, bf16 targets (rejected: no significant gain) · 30000
Run en25-hubl18: EN-25 REPA-teacher arm: HuBERT-large layer 18 (rejected: no significant gain) · 30000
Run en25-hubl24fp32: EN-25 REPA-teacher arm: HuBERT-large layer 24, fp32 targets (rejected: no significant gain) · 30000

Screening runs

Small Tier-1 models: ~80 M parameters, 30k steps, 88.8 hours of speech (31,987 clips, 3,587 voices). They were trained only to choose settings; their quality is far below the planned main model.

Run en55-t06

EN-55 noise-schedule arm: t_mean -0.6 / t_std 0.9 (rejected: worse on echo_dev, seed_dev)

en55-t06 step 30000

echo-dev WER 3.34 % · SIM-o 0.676 · UTMOS 3.09 (GT 4.20) | seed-dev WER 8.01 % · SIM-o 0.448 · UTMOS 2.96 (GT 3.52) · EN-E1: worse on echo_dev, seed_dev

Model: checkpoints/en55-t06/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en55-pc06-t06

EN-55 noise-schedule arm with the EN-29 pc06 prompting: t_mean -0.6 / t_std 0.9 (rejected: worse on seed_dev)

en55-pc06-t06 step 30000

echo-dev WER 2.14 % · SIM-o 0.687 · UTMOS 3.09 (GT 4.20) | seed-dev WER 6.56 % · SIM-o 0.449 · UTMOS 2.99 (GT 3.52) · EN-E1: worse on seed_dev

Model: checkpoints/en55-pc06-t06/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en55-pc06-t08

EN-55 noise-schedule arm with the EN-29 pc06 prompting: t_mean -0.8 / t_std 0.8 (EN-E1: no significant gain; the Tier-1 arm of the main-run (M4) recipe)

en55-pc06-t08 step 30000

echo-dev WER 1.68 % · SIM-o 0.681 · UTMOS 3.04 (GT 4.20) | seed-dev WER 5.30 % · SIM-o 0.440 · UTMOS 2.94 (GT 3.52) · EN-E1: no significant gain

Model: checkpoints/en55-pc06-t08/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en55-pc06-t10

EN-55 noise-schedule arm with the EN-29 pc06 prompting: t_mean -1.0 / t_std 0.8 (rejected: no significant gain)

en55-pc06-t10 step 30000

echo-dev WER 2.07 % · SIM-o 0.663 · UTMOS 2.95 (GT 4.20) | seed-dev WER 5.21 % · SIM-o 0.413 · UTMOS 2.79 (GT 3.52) · EN-E1: no significant gain

Model: checkpoints/en55-pc06-t10/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en55-t08

EN-55 noise-schedule arm: t_mean -0.8 (rejected: worse on echo_dev, seed_dev)

en55-t08 step 30000

echo-dev WER 3.38 % · SIM-o 0.662 · UTMOS 2.93 (GT 4.20) | seed-dev WER 8.01 % · SIM-o 0.435 · UTMOS 2.78 (GT 3.52) · EN-E1: worse on echo_dev, seed_dev

Model: checkpoints/en55-t08/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en29-fixed-pc06-tf04

EN-29 prompting arm: fixed picker, p_cross 0.6, text drop 0.4 (adopted)

en29-fixed-pc06-tf04 step 30000

echo-dev WER 1.24 % · SIM-o 0.649 · UTMOS 2.76 (GT 4.20) | seed-dev WER 5.97 % · SIM-o 0.407 · UTMOS 2.58 (GT 3.52) · EN-E1: adopted: adopt

Model: checkpoints/en29-fixed-pc06-tf04/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en29-fixed

EN-29 prompting arm: fixed cross-prompt picker (rejected: no significant gain)

en29-fixed step 30000

echo-dev WER 1.73 % · SIM-o 0.639 · UTMOS 2.76 (GT 4.20) | seed-dev WER 5.21 % · SIM-o 0.392 · UTMOS 2.65 (GT 3.52) · EN-E1: no significant gain

Model: checkpoints/en29-fixed/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en29-fixed-pc06

EN-29 prompting arm: fixed picker, p_cross 0.6 (adopted)

en29-fixed-pc06 step 30000

echo-dev WER 1.06 % · SIM-o 0.652 · UTMOS 2.71 (GT 4.20) | seed-dev WER 6.30 % · SIM-o 0.424 · UTMOS 2.63 (GT 3.52) · EN-E1: adopted: adopt

Model: checkpoints/en29-fixed-pc06/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en55-repafade

EN-55 arm: REPA faded out after 18k steps (rejected: worse on echo_dev, seed_dev)

en55-repafade step 30000

echo-dev WER 3.15 % · SIM-o 0.628 · UTMOS 2.70 (GT 4.20) | seed-dev WER 7.28 % · SIM-o 0.398 · UTMOS 2.51 (GT 3.52) · EN-E1: worse on echo_dev, seed_dev

Model: checkpoints/en55-repafade/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en-t1-base-s2

baseline, seed 2

en-t1-base-s2 step 30000

echo-dev WER 2.51 % · SIM-o 0.630 · UTMOS 2.70 (GT 4.20) | seed-dev WER 5.84 % · SIM-o 0.392 · UTMOS 2.62 (GT 3.52) · EN-E1: baseline (seed 2); vs the other seed: no significant gain

Model: checkpoints/en-t1-base-s2/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en-t1-base

baseline, seed 1234

en-t1-base step 30000

echo-dev WER 1.96 % · SIM-o 0.631 · UTMOS 2.68 (GT 4.20) | seed-dev WER 5.96 % · SIM-o 0.393 · UTMOS 2.58 (GT 3.52) · EN-E1: baseline (seed 1234)

Model: checkpoints/en-t1-base/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en25-hubl24

EN-25 REPA-teacher arm: HuBERT-large layer 24, bf16 targets (rejected: no significant gain)

en25-hubl24 step 30000

echo-dev WER 1.68 % · SIM-o 0.617 · UTMOS 2.59 (GT 4.20) | seed-dev WER 5.37 % · SIM-o 0.381 · UTMOS 2.45 (GT 3.52) · EN-E1: no significant gain

Model: checkpoints/en25-hubl24/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en25-hubl18

EN-25 REPA-teacher arm: HuBERT-large layer 18 (rejected: no significant gain)

en25-hubl18 step 30000

echo-dev WER 2.03 % · SIM-o 0.607 · UTMOS 2.59 (GT 4.20) | seed-dev WER 5.49 % · SIM-o 0.369 · UTMOS 2.54 (GT 3.52) · EN-E1: no significant gain

Model: checkpoints/en25-hubl18/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Run en25-hubl24fp32

EN-25 REPA-teacher arm: HuBERT-large layer 24, fp32 targets (rejected: no significant gain)

en25-hubl24fp32 step 30000

echo-dev WER 1.75 % · SIM-o 0.614 · UTMOS 2.55 (GT 4.20) | seed-dev WER 5.82 % · SIM-o 0.394 · UTMOS 2.34 (GT 3.52) · EN-E1: no significant gain

Model: checkpoints/en25-hubl24fp32/step_0030000 · trained on echo_en · sampling: 32 Euler steps + sway, joint CFG w 4.0, initial noise 0.9, at most -16 LUFS (peak-limited), duration by the band rule

Item	Category	Sample
`01-short`	short everyday sentence
`02-question`	question
`03-numbers`	numbers, dates, abbreviations
`04-conversational`	medium conversational turn (~10 s)
`05-long`	long passage (~20 s, two sentences)

Licence

apache-2.0 (owner decision 2026-09-29, hobby project): the samples (model outputs) and the prompts, which are clips of held-out voices of SynDataLab-EN/echo-clones-4m-en (apache-2.0 on its card).

Responsible use

The samples are synthetic speech in the prompts' voices and are watermarked. Do not use them, or the model, to impersonate a real person or to deceive.

Downloads last month: 378

Number of rows:

Total file size:

70.2 MB